Activity for Opus 5.5 and GPT-6 Sol/Luna Put Coding-Agent Economics to the Test

Opus 5.5 and GPT-6 Sol/Luna Put Coding-Agent Economics to the Test

September 23, 2026 •

Anthropic’s Opus 5.5 and OpenAI’s GPT-6 Sol/Luna arrived with lower-cost claims and early coding results; a first-party HAProxy migration gives the clearest concrete signal. The brief pairs model economics with practical workflows for effort selection, PR completion, formal verification, and agent tool search.

anthropic.com

SToneX

Simon Willison's Weblog

🔥 TOP SIGNAL

Compare cost per completed change, not token rates alone. Anthropic says Opus 5.5 matches Claude Fable 5.1 for most tasks at 40% lower cost on typical workloads; it is now the medium-effort default in Claude Code and the Claude app for Pro, Max, and Team. OpenAI’s GPT-6 Luna starts at $0.10 per million input tokens and $0.50 per million output tokens.

The strongest coding datapoint is Boris Cherny’s internal HAProxy C-to-Rust test: Opus 5.5 and Fable 5.1 both passed nearly all tests, but Opus 5.5 finished in 9.5 hours versus 12 and cost 51% less. Treat it as a promising first-party result, not an independent bake-off.

⚡ TRY THIS

Sweep effort settings on a real task. Matthew Berman’s read of Anthropic’s FrontierCode comparison says Opus 5.5 at medium effort scored higher than max while costing under $1 versus over $5 per task. Run a representative repo task at both settings and compare test results, elapsed time, and cost before pinning a default.
Give long-running coding work an explicit end state. In Theo’s T3 Code workflow, the prompt supplied a screenshot and thread ID, asked the agent to find root causes, fix them, and file one focused PR. Say whether it should stop at the PR, babysit it until CI is green, or merge once green; Theo’s babysit instructions check only comments and CI newer than the latest push, verify bot findings, and avoid scope creep.
Try an agent-assisted formal-verification pass on risky code. Boris Cherny says a couple of short prompts with Opus 5.5 and Lean produced 16 PRs fixing bugs and race conditions in the Claude Agent SDK; he sometimes combines Lean and TLA+ to probe data flow, concurrency, and state management.
For tool discovery, keep lexical recall broad and rerank candidates with Jev. Treg takes the top 30 lexical matches on any rare query term, asks Jev whether each tool would directly accomplish—or be necessary for—the task, drops scores below 0.4, and prioritizes scores of 0.7 or higher while retaining lexical order and existing success/price adjustments. Its interleaving favored the new pipeline by 8.6:1 points excluding ties—not 9× search accuracy—and the test combines broader retrieval with Jev, so it does not isolate Jev’s contribution. Treg’s write-up.

📡 WHAT SHIPPED

Claude Opus 5.5: Anthropic lists API rates of $4/$20 per million input/output tokens and $0.20 per million cache reads—20% lower input/output prices and 60% cheaper cache reads than Opus 5. It claims 40% lower cost on typical workloads at default settings and over 30% faster output; Anthropic also cautions that benchmark margins are a less reliable guide to real-world differences than before.
GPT-6 Sol and Luna: OpenAI lists Sol at $2/$10 and Luna at $0.10/$0.50 per million input/output tokens. Its DeepSWE 1.1 results report Sol at 68.8% on max effort, within 1.1 points of Claude Fable 5 at xhigh and at about 80% lower task cost; Luna scores 66.6% at max, described as comparable to Opus 5 and Fable 5 at medium effort. These are OpenAI-reported evaluations; validate them in your own harness.
LangChain’s Patch demo turns a Slack feature request into a GitHub PR ready for review. The Managed Deep Agents recipe uses agent.py for the agent/model, instructions.md for the repo and procedure, a Slack channel, GitHub MCP with its token in an environment-variable secret, and define_sandbox for coding and tests.
Simon Willison’s LLM CLI updates:llm 0.36 adds gpt-6-sol and gpt-6-luna, plus a guard for single-turn models; llm-anthropic 0.29 adds Opus 5.5 (llm -m claude-opus-5.5 "prompt goes here"); llm-typesafe 0.1a0 adds Jev support (llm install llm-typesafe).

🎬 GO DEEPER

Matthew Berman on Opus 5.5’s effort economics: The useful segment compares effort settings and cost per task; watch it before assuming max is the best default.

Theo on long-running Fable coding work: A firsthand PR workflow—context, root cause, review-bot feedback, and a defined stop condition.

LangChain’s Patch walkthrough: See the Slack-to-PR agent’s setup, including its instructions file, GitHub access, and sandbox.

Repo — llm-typesafe: A quick way to inspect Jev’s typed yes/no, choice, and scoring calls from the LLM CLI; the release includes install and API-key setup.

Editorial take: Cheaper models widen the options; effort tuning, explicit PR completion criteria, and verification determine whether the savings turn into shippable code.

Opus 5.5 and GPT-6 Sol/Luna Put Coding-Agent Economics to the Test

Research extraction

GPT‑6 Sol and Luna are listed as available in the OpenAI API under gpt-6-sol and gpt-6-luna. The launch post gives API prices per 1 million tokens: Sol $2 input / $10 output and Luna $0.10 input / $0.50 output.

Cached input: The post says cached input-token reads receive a 90% discount. Applying that discount to the listed input rates implies $0.20 per 1 million cached input tokens for Sol and $0.01 for Luna; these effective rates are calculated from the announcement, not separately printed.
Pricing comparison caveat: The listed GPT‑5.6 promotional rates are Sol $4/$20 and Luna $0.20/$1.20 (input/output). The post labels each GPT‑6 model “50% cheaper,” but Luna’s listed output price, $1.20 to $0.50, does not equal a 50% reduction.
Coding: The post says GPT‑6 Sol improves substantially over GPT‑5.6 Sol on FrontierCode and matches Claude Fable 5.1 xhigh at much lower cost, without giving a numerical GPT‑5.6-to-GPT‑6 score change. On DeepSWE v1.1, max-effort Sol scores 68.8%, versus Claude Fable 5 at xhigh’s 69.9%, at approximately 80% lower cost per task; max-effort Luna scores 66.6%, described as comparable to Claude Opus 5 and Fable 5 at medium effort.
Computer use: On OSWorld 2.0 offline, Sol at xhigh scores 60.5% versus Claude Opus 5 at medium effort’s 60.3%, at approximately 80% lower cost per task. Luna at max effort exceeds GPT‑5.6 Sol at medium effort at one tenth of its cost.
Other GPT‑5.6 comparisons: On the post’s internal factuality evaluation, Sol makes about half as many mistakes as its predecessor; at higher effort, Luna matches GPT‑5.6 Sol at about one hundredth its cost. The post cautions that this evaluation uses conversations where users had flagged factual errors and is not representative of typical usage. Both models also show lower rates of misleading claims about their coding work than their GPT‑5.6 counterparts; the post says these evaluations deliberately test challenging situations and do not measure typical-use failure rates.
Availability context: The post says Sol and Luna are available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise, and Edu users, while Free and Go users can access Luna in the desktop app; the models are not yet available in Chat. ChatGPT rollout is described as gradual.

Introducing GPT-6 Sol and Luna | OpenAI

Research extraction

Direct answer: The supplied announcement claims strong agentic-coding results and lower typical-workload costs, publishes reduced token and cache rates, and describes developer-facing speed, security, and API changes. Its CursorBench table score and prose margin claim do not reconcile.

Coding benchmarks: The table reports Opus 5.5 at 66.4% on Terminal-Bench 4.0, 54.4% on FrontierCode v1.1, and 57.8% on CursorBench 4.0; it shows higher scores than the listed comparison models on those coding tests. The announcement cautions that benchmark margins may be a less reliable guide to real-world differences and says its own performance gap versus Fable 5.1 is narrower than the scores suggest. Unless otherwise noted, Claude results use adaptive thinking at max effort; Terminal-Bench 4.0 uses Opus 5.5 at xhigh and GPT-6 Astra at high effort. The footnote gives a ±2.6-point standard error for Opus 5.5 on that benchmark and says the public leaderboard uses the Claude Code harness and five trials per task.
Unreconciled CursorBench claim: The table’s 57.8% for Opus 5.5 and 41.7% for GPT-5.6 Sol imply a 16.1-point difference, whereas the coding narrative says Opus 5.5 beats GPT-5.6 Sol by 11 points; the supplied announcement does not explain the difference.
Task cost and speed claims: Anthropic says Opus 5.5 needs less compute to serve, costs 40% less than Opus 5 on typical workloads at default settings, and generates output more than 30% faster. It attributes the task-cost reduction to both lower per-token cost and fewer tokens per task. A reported example is a 200,000-line codebase audit and fix completed in under three hours, versus over 20 hours for Opus 5, which used 2.5 times as many tokens; in an internal HAProxy C-to-Rust task, Opus 5.5 finished in 9.5 hours versus Fable 5.1’s 12 hours and cost 51% less. For per-task comparisons, the announcement says Opus 5.5 at default effort beats GPT-6 Astra on FrontierCode for roughly one-fifth the cost, matches Astra on Terminal-Bench for about 40% of the cost, and beats GPT-5.6 Sol on CursorBench by 11 points for about one-third the cost.
Token pricing per million tokens: Opus 5.5 is listed at $4 input, $20 output, $0.20 cache reads, and $5 cache writes; the corresponding Opus 5 rates are $5, $25, $0.50, and $6.25. Fast mode is available in Claude Code and Claude Platform, offers up to 2.5× speed, and costs $8 per million input tokens and $40 per million output tokens.
Harness and developer-facing details: The named harness detail is methodological: the Terminal-Bench 4.0 public leaderboard uses the Claude Code harness and five trials per task; Terminal-Bench-Science uses that harness and three trials per task. The announcement does not specify a new harness version or configuration. It describes a classifier that screens each action before execution, an open-source sandbox that security teams can audit, and code review intended to catch vulnerabilities before merge. It also says routine software-development bug-fixing is allowed while most cybersecurity tasks are rerouted to Opus 4.8.
API/model behavior and availability: Preserved thinking prevents API users from editing Claude’s prior context to extract its reasoning; the announcement says it applies to Opus 5.5 and Fable 5.1 for API accounts created on or after August 31, 2026. Opus 5.5 is also no longer available with thinking switched off. The announcement lists claude-opus-5-5 as the Claude Platform model ID and says the model is available across platforms, including AWS, Google Cloud, and Microsoft Azure.

Introducing Claude Opus 5.5

Theo - t3․gg

youtube

For agentic coding comparisons, track total tokens, cost, and tool calls per prompt—not token usage per API request alone: each tool call creates another request, so request-level figures can hide higher end-to-end usage on tool-heavy tasks.
Theo describes an informal, firsthand repo-audit evaluation: have several models find improvement opportunities in a real codebase, then use Fable 5.1 and Astra as LM judges of suggestion quality. On T3 Code, Astra led, Grok 4.7 was close, and Fable 5.1 made five suggestions versus eight for the other models, with valid suggestions but less verification. Grok 4.7’s run cost ranged from roughly peer-level to twice as much at worst, which Theo attributed to it spending extra effort verifying low-value details.
Theo found Grok 4.7 more persistent and inquisitive, but its strengths were task-dependent: in a same-prompt, same-environment front-end/game test, he judged its result poor compared with Astra, and reports one run took nearly 90 minutes.
In a sponsored segment, Theo says his team handles hundreds of PRs a day and now runs 8–30 checks per PR as it adds CI to verify agent work; he argues fast, inexpensive feedback matters at that scale. He describes Blacksmith as replacing GitHub Actions runners with a one-line workflow change, and says demo prompts in its Codesmith tool led him to merge three CI-improvement PRs.

Elon promised this one would be good...

Matthew Berman

youtube

An Anthropic Claude Code team member recommends Opus 5.5 as a daily driver for general software engineering; they used it for an entire personal-site redo, while reserving Fable 5.1 for discrete planning, code review, and security tasks where the cost of a mistake matters more than cost efficiency. Opus 5.5 was available in Claude Code at launch.
Compare cost per completed task, not token price alone: Berman relayed Anthropic’s claim that Opus 5.5 costs 40% less than Opus 5 on typical workloads at default settings, despite about a 20% reduction in listed token prices, attributing the difference to fewer tokens needed to finish tasks. In Berman’s Frontier Code comparison, Opus 5.5 at medium effort scored higher for under $1 per task than max effort at over $5; test effort settings on your own workload rather than assuming max is best.
The Claude Code team has removed much of its system prompt and adjusted tools as models improve. Alongside the release, it introduced plugin evaluations to check whether skills and plugins are current for a new model and help update them so they do not constrain it.
In an early-access demo, Berman’s teammate Alex asked Opus 5.5 to make a Dark Souls-like game in Unreal; Alex said it worked for about 28 hours without Slash Goal and that he now uses natural-language instructions such as “keep working until it’s perfect.” The reported result included six worlds, NPCs, an upgrade system, and roughly 15 bosses—an example of a long-running side-project workflow, not production evidence.

GPT-6 SOL AND LUNA ARE OUT!!!

Latent.Space

latent

Latent.Space reports Xiaomi’s open-weight MiMo-V2.6-Pro and Flash release as natively omnimodal; Xiaomi positions Pro as its most capable model and Flash as the efficiency/cost balance, and says Pro-UltraSpeed is rolling out with up to 20× faster output at the same quality. Artificial Analysis reports Pro at 1.02T total/42B active parameters, with a score of 46 on its Intelligence Index and listed rates of $0.435/M input and $0.87/M output.
Xiaomi’s technical report describes an RL recipe using a fully asynchronous architecture, 1,568 samples per update, up to 1M context length, and 3.5–3.7B tokens per step. It mixes coding, general-agent, visual, and cyber tasks across harnesses; group-relative comparisons provide more varied reward signals for long-horizon tasks and steer training toward shorter paths and fewer tokens per task.
For implementation, Xiaomi links coding recipes, dataset loader, and rewards and composable agent configurations. Xiaomi says the environments and training recipes will be open-sourced, but its 7K+ task datasets had not yet been released.
A transferable orchestration pattern from the roundup is to use specialized decision models for routing, approval gates, trace scoring, tool selection, and low-cost supervision inside larger agent loops, rather than treating them as standalone agents.
The roundup reports Cognition introduced Devin Cloud in Terminal and devin ssh, exposing Devin’s VM through the CLI and enabling handoff between Devin and the user’s machine; it also says GitHub Copilot had teased editable diffs and shows a Sentry-integrated canvas moving from crash report to fix. These are reported product updates/demos, not firsthand workflow accounts.

[AINews] Xiaomi MiMo-V2.6-Pro 1T-A42B: the new top Open Weights model, trained for $3M

Latent Space

youtube

John Platt, Google Fellow and head of applied science at Google Research, described ERA as a specialized Gemini-powered research harness; the paper-era setup discussed used Gemini 2.5, and Platt said it likely would not have worked with Gemini 2.0.

Workflow: A researcher describes a problem, and an agent helps define a scorable objective and creates a Python notebook; supplied papers can guide the initial code, while starter code is optional. ERA then mutates and tests notebook candidates in a tree and can recombine ideas from different candidates.
Orchestration: ERA keeps shared history of attempts and results, pruning it to manage context. Its default is 10 search leaves at a time: larger batches limit cross-learning. It uses an upper-confidence-bound strategy to favor candidates with promising performance and uncertainty, rather than always choosing the current top scorer.
Human oversight and rigor: Platt says people often need to refine the objective when the agent finds loopholes, and recommends hidden holdout sets to guard against overfitting. ERA can run for hours before a person reviews examples and steers the next round.
Firsthand result: In Google’s contrail counterfactual modeling work, the team had been stuck for two years on estimating reflected sunlight; ERA searched confounders and found a simple model that passed synthetic-data tests the team’s earlier attempts had failed.

🔬 Google's AI Scientist Started as an Attempt to Automate Kaggle — John Platt, Google Fellow

Matthew Berman

youtube

Model choice (firsthand use): An Anthropic technical-staff guest says they used Opus 5.5 throughout a personal-site rebuild and sees it as a daily driver; they’d use Fable 5.1 for discrete planning, code review, or security tasks where the cost of error matters more than cost efficiency.
Tune effort; compare cost per completed task: Matthew Berman’s recap reports Opus 5.5 at 66.4 on Terminal Bench 4.0, ahead of Astra at 57.9 and Fable 5.1 at 55.8. He says that on Frontier Code, medium effort scored higher than max while costing under $1 versus over $5 per task, and recommends experimenting with effort settings rather than defaulting to max. He also reports over 30% faster output and 40% lower cost on typical workloads at default settings versus Opus 5, arguing that cost per completed task matters more than token price alone.
Keep the harness aligned with model capability: The Anthropic guest says the team removed much of its system prompt, changed tools, and released plug-in evaluations to check whether skills and plugins suit the newer model and help update them so they do not constrain it. For parallel-agent oversight, Berman says shorter, important-first completion summaries make it easier to switch between threads when managing 10–20 agents.

Anthropic went CRAZY (Opus 5.5)

Theo - t3․gg

youtube

Model choice and effort: Theo reports that Fable 5.1 is less noisy than Astra and can sustain longer runs; he defaults to High, uses X High for especially deep work, and reserves Low/Medium for deliberately limiting a rabbit hole because a failed lower-effort attempt can mean spending more tokens on a rerun. These are his firsthand usage observations.
Prompt for a finished outcome: State where the agent should stop—such as filing a PR, monitoring it until checks are green, or merging once it is ready—and give it conditional paths when the right next step depends on what it finds. For PR babysitting, Theo’s skill says to check only comments and CI results newer than the latest push, verify bot findings against source, fix real issues, and avoid scope creep or filler comments.
Delegate verification across agents: Theo describes Fable as the stronger code writer and Astra as a somewhat better reviewer; he has Fable ask Astra to review or test changes, then revise and retest before reporting back. For computer-use verification, he says Claude is weaker than Codex on macOS and suggests having Fable call Codex for that work.
Gate risky changes with evidence: In a non-critical Lakebed experiment, Theo had Fable assess a risky runtime change, identify what staging lacked, and build confidence measures including shadow mode, structured refresh-failure reasons, synthetic staging traffic, and runtime health counters. He cautions that this kind of autonomous workflow is dangerous on important codebases without strong staging and QA; during the experiment, laptop-driven testing caused network problems and he told the agent to stop.
Keep prompts and configuration lean: Theo recommends removing stale formatting rules from Agent.md/Claude.md and starting with minimal defaults, adding instructions only to address observed problems. If an agent is making unrequested changes, he relays this prompt from Anthropic’s guide: “If, while working or testing, you find pre-existing bugs, performance concerns, or behaviors the task doesn't mention, don't fix, optimize, or extend it in this change unless the requested behavior cannot work without it. Report it as a follow-up in your summary.”

I was using Fable wrong, this is how I fixed it

Simon Willison's Weblog

simonwillison

LLM 0.36 added gpt-6-sol and gpt-6-luna model support. Model-plugin authors can set supports_conversation = False for single-turn models; LLM raises ConversationNotSupported if they receive assistant or tool history, and llm chat rejects them before a session starts. The first plugin using this capability is llm-typesafe—a useful compatibility guard for integrations that route requests to single-turn models.

llm 0.36

Simon Willison's Weblog

simonwillison

Simon Willison now uses GPT-6 Sol and Claude Opus 5.5 as his default models in Codex and Claude Code; he switched his Datasette Agent demo to GPT-6 Luna, which he says seems fast and competent at SQL queries and building HTML and JavaScript for Datasette Apps.
GPT-6 Luna costs $0.10/$0.50 per million input/output tokens versus GPT-5.6 Luna at $0.20/$1.20; GPT-6 Sol costs $2/$10 versus GPT-5.6 Sol at $4/$20. Opus 5.5 costs $4/$20, 20% below earlier Opus pricing, and its cache-read price fell 60%—notable for long agent conversations, where Willison says 90%+ of input tokens are processed at cached prices.
Caution on Opus 5.5 “max”: in Willison’s SVG-generation test, it spent so long reasoning that it hit the 128,000-token output limit without returning a response; the failure repeated, and each run cost $2.56 and took nearly 20 minutes. He suspects max can overthink to the point of breaking, though this test was not a coding-agent benchmark.

Claude Opus 5.5, GPT-6 Sol, GPT-6 Luna, and a new price war

Kent C. Dodds 🐨

Kent C. Dodds says Kody can be useful beyond development once “everything” is wired up . The accompanying ChatGPT example chains a search for an A Christmas Carol audition email, retrieval of its Google Drive sheet music, song identification, Spotify lookup, and practice-playlist creation; it reports finding three songs and says it is building the playlist .

Here's another thing that I do with Kody. When you have everything wired up, everything seems like just a quick and easy prompt away. It'…

Simon Willison's Weblog

simonwillison

Simon Willison’s llm-anthropic 0.29 adds support for Claude Opus 5.5; it can be invoked through the LLM CLI with llm -m claude-opus-5.5 "prompt goes here".

llm-anthropic 0.29

Addy Osmani

Addy Osmani introduced Claude Opus 5.5, claiming it costs 40% less than Opus 5, cache reads are 60% cheaper, and it performs at the level of Claude Fable 5.1 for most tasks. He called it a step up from Opus 5 and a strong model for agentic coding and computer use. In a reply, he also said it writes more naturally and follows provided writing rules more closely than Opus 5.

Introducing Claude Opus 5.5! 40% lower cost than Opus 5 with cache reads 60% cheaper. It performs at the level of Claude Fable 5.1 for mo… Opus 5.5 is a big step up from Opus 5. This is a great model for agentic coding and computer use. Learn more: [https://anthropic.com/clau… Opus 5.5 writes and communicates much more naturally and follows the writing rules you give it more closely. Glad to see these improvemen…

Kent C. Dodds 🐨

@kentcdodds linked an @dabit3 post relaying a Cognition giveaway announcement: it says Devin now offers models including GPT-6 Astra/Sol/Luna, Claude Opus 5.5 and Fable 5.1, SWE-2, Gemini 3.8 Flash, Grok 4.7, Kimi K3, Inkling, DeepSeek V4.1 Flash and GLM-5.3 Flash; it also lists the Fusion Frontier harness for Fable, Astra, Sol and Opus, and cloud agents on Linux, macOS and Windows. SWE-2 was advertised as free until October 15.

Cognition with $200 plans [![Oprah Lol GIF by Amy Poehler's Smart Girls](https://pbs.twimg.com/tweet_video_thumb/HS2C2QCaAAAKNZF.jpg)](ht… Cognition is giving away 50 $200 Devin Max plans to celebrate the new model launches! ⚡ Now available in Devin: • GPT-6 Astra, Sol, Luna …

Ben Tossell

Ben Tossell says Astra generated the device images and built his “50 years of devices” site; he later added a live leaderboard.

50 years of devices. save which you had or wanted Astra image-gen'd all the devices + built the site [https://bentossell.com/devices/](ht… just added a live leaderboard 🥇 game boy colour 🥈 discman 🥉 game boy / playstation / nokia 3310 only 1 person had the pokewalker 💔 ![](ht…

Simon Willison

Simon Willison says GPT-6 Luna is his favorite model for building product features, citing its cost and speed; he describes it as half the price of GPT-5.6 Luna . OpenAI’s launch post describes GPT-6 Sol and Luna as faster, more affordable models carrying forward much of GPT-6 Astra’s strengths, and says their API prices are 50% below GPT-5.6 promotional pricing .

GPT-6 Luna is half the price of 5.6 Luna, which was already an astonishingly cheap model given how capable it is Luna is my favorite mode… Please welcome GPT-6 Sol and GPT-6 Luna to the GPT-6 universe. GPT-6 Sol and Luna build on the advances behind GPT-6 Astra, bringing much…

Ben Tossell

Ben Tossell says he used Astra to generate imagery for a “50 years of devices” site and build the site itself; the project lets visitors save devices they had or wanted.
Ben shared a demo video made with Nilbuild’s /video-demo skill. Nilbuild describes invoking it as /video-demo [description]; it records a web-app demo with zoom-ins and voiceover, and users can request revisions until satisfied. The skill’s repository is https://github.com/nilbuild/video-demo.

50 years of devices. save which you had or wanted Astra image-gen'd all the devices + built the site [https://bentossell.com/devices/](ht… demo video made with this [https://x.com/nilbuild/status/2101247123886845968?s=20](https://x.com/nilbuild/status/2101247123886845968?s=20… Made a skill for video demos /𝚟𝚒𝚍𝚎𝚘-𝚍𝚎𝚖𝚘 [𝚍𝚎𝚜𝚌𝚛𝚒𝚙𝚝𝚒𝚘𝚗] It records a video demo of your web app with zoom-ins and voiceover. You can keep …

Jason Zhou

Jason Zhou says Treg compared Jev with BM25 across 3,000+ endpoints; Treg’s write-up describes Jev in its production search. For agent API/tool discovery, the team kept lexical retrieval but broadened it to the top 30 candidates matching any rare query term, then batch-scored candidates with TypeSafe Jev’s System One model using a Noul yes/no probability question: “Would calling this tool directly accomplish the task, or be a necessary step toward it, on the platform or data source the task requires?” They dropped scores below 0.4, placed scores ≥0.7 in a priority group and scores from 0.4 to below 0.7 in the next group, preserved lexical ordering within groups, and retained existing success-rate and price adjustments. Jev could judge about 30 candidates per request; the reported input cost was about $0.0002 per search at $42 per billion input tokens.

For evaluation, Treg interleaved the two result lists on the same query and credited the variant that contained or ranked higher the tool the agent actually called. The Jev variant won roughly 8.6:1 points excluding ties (rounded to 9:1)—not 9× search accuracy. A call does not establish task completion, and the comparison combines broader retrieval with Jev reranking, so it does not isolate Jev’s contribution.

We compared Jev VS BM25 for matching best tool to agent task Across 3000+ endpoints we have on Treg Result is stunning More technical lea… How Jev Won 9× as Often in treg’s Search

geoff

Geoffrey Huntley says Underclass now supports automatic Codex subscription resets: when a live request finds every enabled ChatGPT/Codex account unavailable, it redeems one banked reset for the account that would otherwise wait longest for natural quota recovery; the feature is enabled by default.
For agent secret handling, Huntley says he added Preflight as a checkpoint because agents might paste .env into the model; the request path is client → preflight → underclass → model. To try it, run nix run github:ghuntley/preflight -- serve and point the harness at :8081 instead of :8080; Preflight repo.

🆕 Underclass now supports automatic codex subscription resets 🤌 When a live request finds every enabled ChatGPT/Codex account unavailable… These agent's can't be trusted to not paste your .env into the model. So I put a checkpoint in front of it. underclass made 20 subscripti…

Alex Albert

Alex Albert says he has been using Opus 5.5 with Blender and that its modeling and vision capabilities let him build an entire world from one prompt; his example is pre-earthquake Market Street, San Francisco, in 1906. The prompt first asks for a research file based on Sanborn maps, a 1906 film, historical photos, and USGS topography, recording each building’s footprint, height, facade material, occupant, sources, and confidence level. It then specifies Blender Python, reusable generators for period buildings and street objects, assembly from the sourced data, no downloaded meshes, textures, or HDRIs, and a 10-second video.

I've been on a Blender kick with Opus 5.5. Its better 3D modeling and vision mean you can build an entire world from a single prompt. His… Prompt in Claude Tag "Recreate Market Street, San Francisco as it stood on April 17, 1906, the afternoon before the earthquake, in Blende…

Introducing Claude Opus 5.5

Open on anthropic.com Open 17 hours ago

anthropic.com Website

1 of 19 highlights

We’re introducing Claude Opus 5.5, the first model in our new Claude 5.5 family. It performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5.

Claude Opus 5.5 is our first release since we called for pacing the frontier (opens in new tab). It was tested before release by external evaluators, including Frontier Design (opens in new tab) and METR (opens in new tab). On our automated behavioral audit, the most comprehensive alignment test we run, Opus 5.5 is the strongest-performing model we’ve tested to date. It also comes with the safeguards we’ve developed for our most capable models.

Here are some of the improvements you can expect from Opus 5.5:

Performance. Opus 5.5 is a major step up from Opus 5. It’s the new leading model, and early testers saw large jumps in performance on their most complex work. One tester completed a 680,000-line code migration in less than a day—work that would have taken an engineering team weeks. It’s good at finding and fixing inefficiencies in software: when we asked it to cut load times across every page of a web app, Opus 5.5 succeeded 39 of 40 times, while Opus 5 made smaller improvements that also altered the app’s behavior. A different tester had several Claude models build a game from a single prompt; Opus 5.5 scored higher than any other model on the strength of its graphics and polish.

Safety. Opus 5.5 achieves the best scores of any model to date on our automated behavioral audit, our alignment suite that tests Claude across thousands of simulated scenarios. It is much less likely than recent models to take hard-to-reverse actions or act outside the boundaries it’s been given, and it’s more resistant than Opus 5 to prompt injection. We’ve also broadened our alignment testing to cover longer tasks, impossible tasks, and scenarios modeled on real incidents, though it still has limits. Full details of our evaluation are available in the Opus 5.5 System Card (opens in new tab).

Because Opus 5.5 is comparable to Claude Mythos 5.1 in biology and cybersecurity, we’re deploying it with safeguards similar to those on Claude Fable 5.1. Vetted organizations can apply today to our Life Sciences Verification Program (opens in new tab) to use Opus 5.5 for biology research. In the coming weeks we will also be expanding access to our Cyber Verification Program (opens in new tab), and verified cybersecurity practitioners will be able to use Opus 5.5 for their work.

Cost and speed. Opus 5.5 requires less compute to serve than Opus 5, and its pricing reflects that. Our tests show that at default settings it will cost 40% less than Opus 5 on typical workloads. Input and output tokens are $4 and $20 per million, 20% less than Opus 5. Cache reads (which make up the majority of agentic and coding work costs) are $0.20 per million tokens, 60% less than Opus 5. Opus 5.5 also generates output more than 30% faster than Opus 5.

In addition to the price drop, we’re increasing five-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans. We’re also providing subscription users a rate limit reset, which you can now save and use whenever you choose.

Communication. Opus 5.5 communicates more naturally than prior models. Early testers found its writing clearer and easier to follow, which addresses some of the common feedback we heard about Opus 5. It puts the most important information up front, and its style makes it a better work partner over long sessions. As one early tester put it, “it writes the way I do.” In our own use, this has made Opus 5.5’s work easier to follow and check—which is a safety benefit as well as a practical one.

Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks, with many of the same improvements to performance, efficiency, and safety.

Performance and cost-effectiveness

On our benchmarks, Claude Opus 5.5 leads in agentic coding, computer use, and knowledge work. That said, at these levels of capability we’ve found that benchmark margins have become a less reliable guide to real-world differences. In our own use, the gap between Opus 5.5 and Claude Fable 5.1 is narrower than these scores suggest.

	Opus 5.5	Fable 5.1	Opus 5	GPT-6 Astra	GPT-5.6 Sol
Agentic codingTerminal-Bench 4.0¹	66.4%	55.8%	52.3%	57.9%	37.3%
Agentic codingFrontierCode v1.1 (Main)	54.4%	50.3%	48.0%	53.3%	47.5%
Agentic codingCursorBench 4.0	57.8%	51.8%	46.6%	—	41.7%
Knowledge workGDPval-AA v2.1	1846	1735	1708	1542	1588
Business workflowsAutomationBench²	40.0%	31.4%	26.9%	41.4%	28.8%
Multidisciplinary reasoningHumanity's Last Exam	67.7%with tools	65.6%with tools	63.6%with tools	57.2%with tools	—
Agentic scientific researchTerminal-Bench-Science 0.1³	58.7%	52.6%	29.0%	64.6%	22.4%
Computer useOSWorld 2.0	81.8%partial	80.7%partial	74.0%partial	—	—
Visual chart recognitionChartography	89.0%with tools	88.4%with tools	83.4%with tools	—	—

Unless otherwise noted, all Claude Opus 5.5 results use adaptive thinking at max effort. Terminal-Bench 4.0 results are reported for Claude Opus 5.5 at xhigh effort and GPT-6 Astra at high effort, as reported by OpenAI; these represent each model’s highest score. Claude Opus 5.5 was evaluated with its production safeguards enabled. When they intervened, cybersecurity tasks were completed by Claude Opus 4.8, and biology and frontier LLM development tasks were completed by Claude Opus 5. This likely reduces Claude Opus 5.5’s performance on these benchmarks.

[1] Terminal-Bench 4.0: The standard error is ±2.6 pts for Claude Opus 5.5 and ±1.6–2 pts for the other Claude models. The public leaderboard (5 trials/task, Claude Code harness) reports Claude Opus 5 at 51.8%; our setup reproduces it at 52.3%, within noise. GPT-6 Astra and GPT-5.6 Sol figures are as reported by OpenAI.

[2] AutomationBench: AutomationBench results were run and reported by Zapier. These runs were performed without fallback models, so safeguard interventions were considered failures—this resulted in a lower score than Claude Opus 5.5 would achieve in practice. Claude Opus 5.5 results come from Zapier’s own evaluation during early access. Results for Opus 5, GPT-5.6 Sol, and GPT-6 Astra come from Zapier’s public leaderboard.

[3] Terminal-Bench-Science 0.1: The standard error is ±3.5–5 pts per model. The public leaderboard (3 trials/task, Claude Code harness) reports Claude Opus 5 at 30.0%; our setup reproduces it at 29.0%, within noise. The GPT-6 Astra figure is as reported by OpenAI.

Where Opus 5.5’s advantage is very clear is efficiency. It costs less per token than Opus 5 and uses fewer tokens per task, which nets out to a 40% drop in costs.

Pricing

Prices per 1M tokens	Claude Opus 5.5	Claude Opus 5
Cache reads	$0.20	$0.50
Input tokens	$4	$5
Output tokens	$20	$25
Cache writes	$5	$6.25

Fast mode for Opus 5.5 is also available in Claude Code and the Claude Platform with up to 2.5x speed. It costs $8 per million input tokens and $40 per million output tokens.

Coding

Opus 5.5 is particularly good at long and sprawling jobs like codebase-wide migrations and audits. An early tester used it to audit and fix a 200,000-line codebase in under three hours, where Opus 5 took over 20 hours and used 2.5x as many tokens. In an internal test, we asked Opus 5.5 and Fable 5.1 to translate HAProxy, widely used software that balances web traffic loads across servers, from C into Rust. Both rewrites passed nearly all of HAProxy’s own regression tests, but Opus 5.5 finished in 9.5 hours compared to 12 for Fable 5.1, and cost 51% less.

Opus 5.5 delivers frontier results on agentic coding at a fraction of the cost. At its default effort level on FrontierCode, it beats GPT-6 Astra at roughly 20% of the cost per task. On Terminal Bench 4.0, it matches Astra for about 40% of the cost, while on CursorBench it beats GPT-5.6 Sol by 11 points for about a third of the cost.

Terminal-Bench 4.0Accuracy vs Cost

Terminal-Bench 4.0 measures how well a model can complete complex, multi-step professional tasks within a command line interface. Opus 5.5 at default effort beats Opus 5 at max effort for about a fifth of the cost. It matches GPT-6 Astra at about 40% of the cost.

Our early testers reported similar efficiency and intelligence gains:

The most secure coding agent

Enterprises that use agents within their systems need to know that those agents are operating as intended, particularly when they run autonomously for many hours. Opus 5.5 has a classifier that screens every action before it runs, an open-source sandbox that security teams can audit, and code review that catches vulnerabilities before they merge.

The model itself also has stronger defenses. On prompt injection attacks, it matches or beats Opus 5 in every setting we tested, including coding, tool use, computer use, and web browsing. On a benchmark run by the AI security firm Gray Swan, Opus 5.5 ties Fable 5.1 for the lowest prompt injection success rate of any model tested.

Knowledge work

Opus 5.5 is a reliable and adept researcher. In one internal test, we asked Opus 5.5, Fable 5.1, and Opus 5 to write a report on a company’s quarterly performance using only the information it could find on a copy of the web where the earnings release was hard to locate. An automated grader checked every figure and quote against sources. Across different effort settings, 16 out of 18 of Opus 5.5’s reports cleared our quality bar, where any invented figure or quote would have failed. Neither Fable 5.1 nor Opus 5 cleared that bar in any attempt.

It’s also strong in financial analysis and business work. Walleye Capital, an investment firm and early tester, reported that Opus 5.5 largely solved their evaluation suite on its lowest setting; on higher settings, it performed even better, noticing an error in their evaluation instructions and correcting for it. No other model had caught this error before.

In another test, we tasked both Opus 5.5 and Opus 5 with analyzing a proposed merger between two fictional HR software companies. Each built a financial model in Excel, then turned it into an executive presentation on whether the deal made sense at its price. Both models reached the same conclusions about the deal, but Opus 5.5’s model was more thorough and its presentation easier to read, while Opus 5’s had minor errors. Opus 5.5 finished in 63 minutes compared to 93 for Opus 5, and cost 50% less to produce.

On knowledge work evaluations, Opus 5.5 outperforms other models while also using fewer tokens. On GDPval-AA v2.1, a test of real-world work across 44 occupations, Opus 5.5 scores 1846 Elo, ahead of Fable 5.1 and Opus 5. At default effort (medium), Opus 5.5 beats GPT-6 Astra at max effort for about a fifth of the cost per task. It likewise outperformed other models on benchmarks measuring business workflows and large-scale data collection.

GDPval-AA v2.1Elo vs Cost

Artificial Analysis’s GDPval-AA v2.1 evaluates agents on real-world professional work across 44 occupations. At max effort, Opus 5.5 scores 1846 Elo, where Fable 5.1 scores 1735 and Opus 5 scores 1708. At default effort (medium), Opus 5.5 beats GPT-6 Astra at max effort for about a fifth of the cost per task.

Our customers have reported similar results. Here’s what they told us about working with the model:

Communication

We’ve made major improvements to the way Opus 5.5 writes and communicates, one of the most common areas of feedback we heard about Opus 5. Its messages are much easier to understand at a glance, which testers said helped during long working sessions. It puts the most important information up front, is less likely to use jargon or idiosyncratic phrases, and follows the writing rules you give it. We find that this makes Opus 5.5 a noticeably better collaborator. Here’s a side-by-side comparison of the two models:

Claude Opus 5

Please explain the issue to me.

What I found

The extra drop isn’t the free tier — it’s a regression in 0552feb (“extract billing period boundaries into periods.py”, labelled “No behaviour change”).

The bug

aggregate.monthly_usage used to do a half-open interval:

start, end = month_start(year, month), next_month_start(year, month)
if start <= e.ts < end:

It now does:

if period.contains(e.ts)

and in periods.py:

@property
def last_day(self) -> datetime:
    return next_month_start(...) - timedelta(days=1)   # 2026-08-31 00:00:00 UTC

def contains(self, ts):
    return self.first_day <= ts <= self.last_day

last_day is a datetime at midnight, not a whole day. Subtracting one day from Sep 1 00:00 gives Aug 31 00:00, and the comparison is inclusive, so the period effectively ends at the first instant of the 31st. Every event timestamped after midnight on the last day of the month is silently dropped — roughly one day of usage per account, per month.

The new test_periods.py doesn’t catch it because it only checks last_day.date() (which is correctly 2026-08-31) and tests contains at Aug 1 00:00, Aug 15 12:00, Jul 31 23:59 and Sep 1 00:00 — it never probes anywhere inside the 31st.

Claude Opus 5.5

Please explain the issue to me.

The extra drop is a bug in the billing refactor

The free-tier change accounts for only $1.50 of acme’s August drop. The other $9.92 comes from a bug in commit 0552feb. That commit was labelled “No behaviour change,” but it stops counting usage from the last day of the month.

What changed

Before the merge, aggregate.py used a half-open interval:

start <= e.ts < next_month_start(year, month)    # [Aug 1 00:00, Sep 1 00:00)

The refactor moved this into Period.contains and made the interval closed on the wrong boundary:

last_day = next_month_start(...) - timedelta(days=1)    # = 2026-08-31T00:00:00Z
return self.first_day <= ts <= self.last_day

last_day is a timestamp at midnight, not a whole day. So every event after 2026-08-31T00:00:00Z is excluded. The September period starts Sep 1, so it doesn’t pick these events up either. The usage isn’t moved to another month; it’s never billed at all.

Our customers’ feedback supports these findings:

Safety

Pacing the frontier

Last week, our CEO, Dario Amodei, argued that AI progress should be paced (opens in new tab) so that safety practices stay ahead of model capabilities. Pacing is an approach to keeping AI safe, remaining competitive with China, and realizing AI’s benefits, particularly in areas like biology and medicine.

We largely understand the risks today’s models present and are well equipped to manage them. However, more serious risks could emerge quickly as capabilities improve, and we need to prepare for them now. For that reason, our safety work takes place on two time horizons at once:

Safety practices for current models. The current generation of models relies on an established set of practices: extensive alignment testing, pre-release evaluation by outside organizations such as METR and Frontier Design, and safeguards matched to each model’s capabilities in high-risk areas like cybersecurity and biology. We refine these practices with each release. We believe they are appropriate to the worst risks today’s models present, and that they give us a broad, though not perfect, picture of the range of serious risks.

Additionally, we track our ability to train and evaluate aligned models, and we report on both our public and internal models in the risk reports we publish under our Responsible Scaling Policy (opens in new tab), our voluntary framework for managing catastrophic risks from advanced AI systems.

Preparing for future models. We’re preparing our training and evaluation processes in anticipation of more advanced models. We’re tightening how we filter the environments used in reinforcement learning, since flawed environments (opens in new tab) are a major source of misaligned behavior. Additionally, we’re improving our alignment rewards and developing automated processes for producing new, diverse scenarios for safety training. And we are strengthening our security and monitoring (opens in new tab), including a focused effort to improve interpretability-based monitoring and evaluation. We hope such techniques will help reduce our reliance on auditing a model’s chain-of-thought, or the reasoning it writes out while it works.

Models with greater capabilities—such as those that can fully automate the work of AI research itself—require a higher safety standard still. Our calls for pacing were based in large part on our expectation that such models could be trained soon. For these models, we do not assume the measures described above will meet that safety standard on their own. As AI becomes more capable, public policy should play a larger role in making sure the systems people rely on are safe. That capacity takes time to build, and we’ve started to put the infrastructure in place to support it, as described in “ We Must Pace the Frontier” (opens in new tab) and our recent announcement with Accenture (opens in new tab); we expect to share more details on these efforts soon. We will also continue to contribute to policy discussions with government and industry, including on approaches to regulation and international coordination.

Alignment

On our primary evaluation suite, an automated behavioral audit that assesses Claude across nearly 2,000 scenarios, Opus 5.5 scored better than any recent Claude model on nearly every measure of misaligned behavior. It’s also our strongest model on most measures of honesty.

In particular, Opus 5.5 improves over previous models on several of the behaviors that contributed to recent cybersecurity incidents (opens in new tab), including biased or motivated reasoning, attempting to escape a sandbox, and taking harmful actions after concluding it was in a simulated environment. In a new evaluation designed to test a model’s propensity to cross containment boundaries, Opus 5.5 attempted to circumvent boundaries around 85% less often than Opus 5 or Claude Mythos 5.1, and every attempt it made was low severity and self-reported. For teams running Claude unattended across their codebases and systems, this is just as important as raw capability.

However, as we described in our recent alignment assessment (opens in new tab), building evaluations that reliably catch every failure prior to deployment remains an unsolved problem. We see signs that Opus 5.5 often suspects it is being evaluated, which challenges our ability to assess how it will act in the vast variety of real-world settings it is deployed in. As these settings expand and model capabilities increase, we expect this challenge to grow, unless we make progress on interpretability. Although we are confident that Opus 5.5 shows broad improvements in the areas we are able to measure, we pair our own alignment work with the safeguards described below.

Safeguards

As our models grow more powerful, stricter safeguards are one way we prevent new capabilities from becoming tools for misuse. Opus 5.5 is the first Opus model to launch with a similar class of safeguards to Fable 5.1 on cybersecurity, biology, and distillation, all of which fall back to another model transparently.

Cybersecurity. Because Opus 5.5 has extremely strong cyber capabilities, we’re applying cybersecurity safeguards to Opus 5.5 that are similar to Fable 5.1’s. Users will be able to identify and fix bugs in their code as part of the routine software development lifecycle, but most cybersecurity tasks will be re-routed to Opus 4.8.

For cyberdefenders, we’ll soon be expanding our Cyber Verification Program (opens in new tab) to include Opus 5.5. The new program will include three tiers for increasingly permissive trusted access, including access to Claude Mythos models. Claude Security (opens in new tab) is already available with access to Claude Mythos 5.1.

Biology. Opus 5.5 is highly capable in biology, exceeding Opus 5 and matching or beating Claude Mythos 5.1 across many areas of work. For example, Opus 5.5 achieved improvements on a long-horizon molecular prediction and design evaluation conducted in collaboration with Dyno Therapeutics, and expert red-teamers rated its scientific novelty as comparable to the best model they had tested.

For this reason, Opus 5.5 uses the same biology safeguards as Fable 5.1. To use Opus 5.5 for research and development work impeded by these safeguards, users can apply to our new Life Sciences Verification Program (opens in new tab), which gives vetted organizations like academic labs, startups, and pharmaceutical companies access to safeguards designed for the full breadth of biology-related work. Interested organizations can apply here (opens in new tab).

Distillation

Distillation attacks, in which attackers use thousands of fake accounts to extract a model’s capabilities at industrial scale, create safety and national security risks. Distillation allows bad actors to create highly capable models without the safeguards we build into Claude. Our September 2026 threat intelligence report (opens in new tab) details the illicit distillation activity we’ve detected and disrupted so far.

Opus 5.5 is launching with preserved thinking, the anti-distillation safeguard we introduced with Fable 5.1. It stops API users from editing Claude’s prior context in an attempt to extract Claude’s reasoning. It applies to Fable 5.1 and Opus 5.5 for API accounts created on or after August 31, 2026. Our Help Center article (opens in new tab) explains the change, and our preserved thinking docs (opens in new tab) show how to test and update your integrations.

Data retention and compliance

Like previous Opus models, Opus 5.5 is available with zero data retention.

As with Fable 5.1, Opus 5.5 comes with our watermarking measures to comply with the EU AI Act, discussed here (opens in new tab). It is also no longer available with “thinking” mode switched off, as we describe here (opens in new tab).

Availability

Claude Opus 5.5 is now available on all platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure. On the Claude Platform, developers can get started (opens in new tab) with claude-opus-5-5.

Insights

Research extraction

Coding benchmarks: The table reports Opus 5.5 at 66.4% on Terminal-Bench 4.0, 54.4% on FrontierCode v1.1, and 57.8% on CursorBench 4.0; it shows higher scores than the listed comparison models on those coding tests. The announcement cautions that benchmark margins may be a less reliable guide to real-world differences and says its own performance gap versus Fable 5.1 is narrower than the scores suggest. Unless otherwise noted, Claude results use adaptive thinking at max effort; Terminal-Bench 4.0 uses Opus 5.5 at xhigh and GPT-6 Astra at high effort. The footnote gives a ±2.6-point standard error for Opus 5.5 on that benchmark and says the public leaderboard uses the Claude Code harness and five trials per task.
Unreconciled CursorBench claim: The table’s 57.8% for Opus 5.5 and 41.7% for GPT-5.6 Sol imply a 16.1-point difference, whereas the coding narrative says Opus 5.5 beats GPT-5.6 Sol by 11 points; the supplied announcement does not explain the difference.
Task cost and speed claims: Anthropic says Opus 5.5 needs less compute to serve, costs 40% less than Opus 5 on typical workloads at default settings, and generates output more than 30% faster. It attributes the task-cost reduction to both lower per-token cost and fewer tokens per task. A reported example is a 200,000-line codebase audit and fix completed in under three hours, versus over 20 hours for Opus 5, which used 2.5 times as many tokens; in an internal HAProxy C-to-Rust task, Opus 5.5 finished in 9.5 hours versus Fable 5.1’s 12 hours and cost 51% less. For per-task comparisons, the announcement says Opus 5.5 at default effort beats GPT-6 Astra on FrontierCode for roughly one-fifth the cost, matches Astra on Terminal-Bench for about 40% of the cost, and beats GPT-5.6 Sol on CursorBench by 11 points for about one-third the cost.
Token pricing per million tokens: Opus 5.5 is listed at $4 input, $20 output, $0.20 cache reads, and $5 cache writes; the corresponding Opus 5 rates are $5, $25, $0.50, and $6.25. Fast mode is available in Claude Code and Claude Platform, offers up to 2.5× speed, and costs $8 per million input tokens and $40 per million output tokens.
Harness and developer-facing details: The named harness detail is methodological: the Terminal-Bench 4.0 public leaderboard uses the Claude Code harness and five trials per task; Terminal-Bench-Science uses that harness and three trials per task. The announcement does not specify a new harness version or configuration. It describes a classifier that screens each action before execution, an open-source sandbox that security teams can audit, and code review intended to catch vulnerabilities before merge. It also says routine software-development bug-fixing is allowed while most cybersecurity tasks are rerouted to Opus 4.8.
API/model behavior and availability: Preserved thinking prevents API users from editing Claude’s prior context to extract its reasoning; the announcement says it applies to Opus 5.5 and Fable 5.1 for API accounts created on or after August 31, 2026. Opus 5.5 is also no longer available with thinking switched off. The announcement lists claude-opus-5-5 as the Claude Platform model ID and says the model is available across platforms, including AWS, Google Cloud, and Microsoft Azure.

Introducing Claude Opus 5.5