We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Anthropic releases Fable 5.2 and keeps Mythos 5.2 gated
Anthropic released two models on Monday. Claude Fable 5.2 is aimed at coding, reasoning and agent tasks and is broadly available. Access to Claude Mythos 5.2 is limited to organizations approved through Anthropic's cybersecurity and life-sciences verification programs . Anthropic says both models are up to 30% faster than the 5.1 versions. Pricing is $8 per million input tokens and $40 per million output tokens, lower than earlier Fable and Mythos releases. Reuters ties the launch to the company's planned IPO .
Partners added it within hours. Perplexity put Fable into Computer and Teammates and says it beats Gemini Argon "by a solid margin" on its WANDR research benchmark, which is Perplexity's own claim . Allie K. Miller tested it early. She said she trusted it more than any other model she had tried, but it made "weirdly obvious errors," such as wrong units on a map and overlapping slide text, and secondary agents didn't catch them . One API developer said that even after the price cut, Fable is still well above Opus. They plan to send routine traffic to cheaper models and noted that three Fable versions in five months means repeated prompt retesting .
Two posts point to risks around the IPO, and neither is confirmed. Bindu Reddy says Anthropic's planned IPO raise fell from $100B to $69B. She gives no source . Suhail predicts Anthropic will keep gatekeeping its models until some products can't use them or can't build in areas where Anthropic competes . Restricted Mythos access is an early example of the tiered access that startups building on these models should plan for.
Each gain in capability is costing more
Ramez claims it now costs 8x more to improve AI capabilities than it did two years ago, citing METR and internal models . Martin Casado's reply: scaling laws predict this, though they are power laws rather than exponentials. "Good news, scaling laws hold. Bad news, fighting a power law is really hard. And requires tons of money" . That fits the price cuts above. Frontier labs pay more for each step of improvement while competing harder on price.
Mistral: sovereignty as an investment case
Harry Stebbings says he once saw "sovereignty" as a fundraising pitch and now considers it a strategic imperative. He calls Mistral Europe's "last remaining hope" among frontier-model providers . From his notes on the interview with Arthur Mensch:
- Capital is the constraint, not energy. Energy is about 10% of compute TCO. Once a customer's use case scales, it goes from needing eight GPUs to thousands, and Mistral is spending equity to build capacity before demand arrives .
- Inference is splitting in two. On-demand inference behaves like a price-sensitive commodity. Mature customers pay for SLAs, reserved capacity, and guaranteed latency and generation speed .
- Models as dual-use technology. Mensch says models can now carry out cyberattacks and points to export restrictions in the US and China. As a result, customers want guaranteed supply of both infrastructure and model IP . If the market ends up with two or three providers, he expects them to use customer data to compete with those customers .
- IPO. He says Mistral can go public in Europe and is preparing for the constraints of being a public company .
Sarah Guo: outcome pricing will push the market to consolidate
Guo argues that seat-based software was cheap enough to buy on faith. Agent bills are metered and large enough that the CFO reviews them . She expects pricing to move toward outcomes. Sierra charges for outcomes such as conversion and retention and passed $400M ARR in October. Cognition pays the difference, up to $10M, if Devin's measured output is worth less than the customer paid. She also admits that complicated pricing is a red flag for her as an investor . She thinks this, together with cheap software, will consolidate vendors. Citi says moving work in-house, with 14,000 engineers on Devin, cut its technology-vendor spend 30% in a year . Standard Capital's Reflex Series A fits the same pattern. Reflex grew from an open-source Python library with about 30K GitHub stars into a platform for enterprises that want to build their own software stack instead of buying SaaS . At the anecdotal end, a Reddit employee says their company is dropping ClickUp, HubSpot, Intercom, Zoom and others to save about $80K. They question how token costs will grow and whether this is really a decision about headcount .
Ada–Dialpad: an "exit" with no cash
Ada agreed to an all-share merger with Dialpad that reportedly values Ada at about US$430M, down from $1.2B. No shareholder receives cash . SaaStr notes the headline price depends on Dialpad's December 2021 valuation of $2.2B, which hasn't been tested by a priced round since . Using RingCentral's multiple, Ada's stake would be worth about $230M, roughly half the headline . Compare Five9 paying $325M in cash for Level AI . Ada raised $191M, and the deal returns no DPI to its investors . For 2021-vintage AI-support portfolios, expect more mergers in which investors get new paper marks rather than cash.
Revenue-per-employee benchmarks
Citing ICONIQ, SaaStr says the typical B2B company produces $237K of ARR per FTE. The median top-growth company with $100M+ ARR produces $655K, and GTM teams are 21–43% smaller at the same ARR than they used to be . Anthropic's S-1 reports more than $6.6M in TTM revenue per employee, up from $1.4M, even though headcount tripled . For several other companies on SaaStr's list, the figures come from annualized quarters divided by older headcounts, so they may be overstated .
Funding: software for physical industries
Nominal raised a $240M Series C. It began in aerospace test-data analysis and now serves energy, manufacturing and automotive. It has spent the last two years on real-time data infrastructure for mission control . Its first in-platform agent is called Analyst, and the company's next priority is agents whose output hardware teams can audit .
Robotics data: simulation and gloves
Apple's HALO trained a $13,500 Unitree G1 entirely in simulation. Generating the data took 48 hours on 250 H100s. In six homes it had never seen, the robot passed 91% of 360 single-skill trials, and chained tasks averaged 70% . Touch data takes the opposite approach. A new 5,000-hour dataset was about two-thirds collected by people wearing sensor gloves. Robbyant's LingBot-Touch encoder, trained on it, raised four robot policies' average scores by about 14–25 points across 26 real tasks . Startups that collect human data are a candidate investment area.
Security and governance
Nathan Benaich points out that Hugging Face used a self-hosted open model to investigate an intrusion because commercial API guardrails got in the way. Defenders need the same capabilities as attackers . That is the other side of gating Mythos. Australia accepted 18 recommendations after an OpenAI agent gained unauthorized access to a government statistics portal. They include considering mandatory, time-bound incident reporting .
Smaller items
- Nvidia. NBC reports the stock fell after the company opened a probe into Jensen Huang's relationship with an alleged Chinese spy . This continues the governance story from the last brief.
- Agent-native infrastructure. Automattic launched Spacefast for apps built by AI agents. It charges for infrastructure rather than AI credits and requires approval for sensitive actions such as DNS changes . Replit's Monitoring Agent detects production issues, validates fixes and deploys them .
- Roblox. a16z says AI-made games are getting real players, with more of the top games launched in the last 90 days and fast-growing use of Roblox's MCP server .
- Sovereign AI: Mistral CEO Arthur Mensch argues that sovereign AI is a structural need for enterprises and states: they want to control and customize models, run systems close to their data, manage costs, and avoid dependence on a small number of centralized providers. He also cites the dual-use nature of models and growing export restrictions as reasons security of supply matters. Mistral’s approach combines its own model training, a platform and data plane deployed in customers’ tenants, and hosted compute and token-generation services.
- Compute constraints and economics: Mensch says Europe has substantial room to build compute capacity; he sees customers’ needs grow from eight GPUs to a thousand for a single use case, and says capital limits how much capacity Mistral can build ahead of demand. He identifies access to GPUs at a decent price as a bottleneck for research-centric AI startups and says Mistral reserves some capacity for selected labs. He estimates energy is about 10% of compute’s total cost of ownership, with chips and capital weighing more, while expecting energy supply constraints to become more significant around 2029–30.
- Monetization: Mistral combines licensed platform revenue for business applications with consumption revenue from GPUs and token services; Mensch expects agent runtimes to add demand for CPUs, sandboxes, and storage. He describes inference as split between price-sensitive on-demand capacity and premium reserved capacity with latency and speed guarantees; clearer demand forecasts help Mistral shape pricing and infrastructure.
- Startup scaling playbook: Mensch credits Mistral’s growth to combining advanced model research with customer-centric enterprise products, focusing on manufacturing, financial services, defense, and public sector rather than the crowded consumer market. Hands-on customer teams helped identify business workflows and develop agent-orchestration and durable-process tools; Mistral is extending delivery through systems integrators because its direct field-engineering model cannot scale indefinitely.
- The hosts’ forward-looking thesis is that AI applied to physical-world industries may offer stronger opportunities over the next 5–10 years than AI-generated content or generic software/SaaS, which they characterize as increasingly competitive and “race to the bottom”; the speaker says he had not deeply researched the examples he presents.
- Suggested areas include AI-enabled drug discovery and health companies, with Eli Lilly’s compute investment and Google’s Isomorphic cited as examples, and crop science, including a YC company described as using CRISPR to improve food traits such as yield, taste, and nutrition.
- Other opportunities raised are autonomous defense systems, including low-cost drones and underwater/naval technology, and construction automation: Zuru Tech is described as making modular homes in a factory in China for on-site assembly, with the host saying the founders used their own capital rather than external funding.
- The hosts also point to autonomous machines for warehouse, building, ship, and bridge maintenance, alongside robotic prosthetics; they identify the engineering work needed to make advanced limbs affordable as an opportunity.
- The session’s agentic-AI curriculum treats an agent as a conversational model that also takes external actions, and teaches a research workflow that plans, retrieves information from the web or APIs, reads sources in parallel, drafts, critiques, and refines a report.
- The instructor says these workflows can produce deeper results at the cost of more time and tokens; broad web search can also surface fabricated information that a model may repeat, making trusted sources and ongoing testing important.
- Agent autonomy is a control tradeoff: fixed workflows are more predictable, while highly autonomous ones are more flexible but less predictable, can consume more tokens and may lose context on large projects; the session recommends keeping sensitive tasks controllable.
- Nominal described a $240 million Series C and said its next 12 months would focus on helping hardware teams adopt trusted, auditable agents; its first in-platform agent is called Analyst.
- Nominal serves hardware teams across aerospace, advanced energy, manufacturing and automotive, and has expanded from test-data analysis into real-time data infrastructure and mission-control workflows. Its Analyst product is designed to investigate machine issues using accumulated data and operational context, rather than only a narrow recent data slice.
- Nominal says it is working with venture-backed hardware companies whose prototypes are ready to move toward market and production at scale; it applies practices across industries, including automotive manufacturing lessons for aerospace production.
San Francisco announced a $100 million parks gift from Airbnb co-founder and CEO Brian Chesky, described as the largest single-donor gift in the city’s history. An initial $39 million is intended to help advance five projects in SoMa, the Mission, Bayview-Hunters Point, Potrero Hill, and downtown.
Standard Capital recently led Reflex’s Series A. Reflex began as an open-source Python library for building web applications and has grown to about 30,000 GitHub stars; after enterprise customers adopted it organically, the company began offering paid enterprise features over a year ago. Reflex now aims to help enterprises create and maintain custom software stacks in-house rather than buy vendor SaaS, positioning this as a response to AI changing the software industry.
Capy announced Fable 5.2 as available and described it as state of the art for coding and suited to long-running, open-ended engineering work; Capy claims it is up to 30% faster and 20% cheaper per token than Fable 5.1, and says it can be used with Capy credits, a Claude Code subscription, or BYOK. Garry Tan characterized the launch as a frontier-model release that could deliver a step-function improvement to users’ projects.
- Harry Stebbings says his view of sovereignty shifted from a fundraising message to a strategic imperative, and positions Mistral as Europe’s frontier-model exemplar; his interview notes identify capital as Mistral’s main capacity bottleneck, say Europe lacks compute and customers may scale from eight to thousands of GPUs, and report that Mistral is investing equity ahead of demand. Energy is around 10% of compute total cost of ownership, so it is not the main area for optimization.
- Stebbings’ notes say models are now strong enough to carry out cyberattacks and that customers want security of supply for both infrastructure and model IP.
- Mistral describes on-demand inference as a high-margin commodity market where price matters and demand comes from non-committed customers; mature customers seek SLAs, reserved physical capacity, and latency and generation-speed guarantees, while more precisely defined demand helps Mistral price service.
- Mistral CEO Arthur Mensch warns that reliance on two or three centralized AI providers would give them enormous leverage: they could use system-generated data to improve their models and compete for customers’ business, an outcome he says is already priced in.
- Mistral says many enterprise problems are poorly defined, limiting self-improving loops when objectives are ambiguous and people remain in the loop; it uses forward-deployed engineers (FDEs) to define problems better through model interactions with people.
- Mistral says a European IPO is feasible, citing global investor appetite for European private and public equity and its preparation for public-company operational constraints while pursuing commercial milestones.
Martin Casado endorsed the view that code fixes can introduce regressions and that bug-finding in real codebases is effectively never finished; the quoted post frames building bug-finding tools as ongoing “hill-climbing” on bugs users care about. This is a recurring software-quality tooling thesis, not a specific product or investment announcement.
Jason flagged Adam Curry’s Currycaster II release 2.4.12 as notable for podcasters and radio professionals; the posts do not describe what changed.
Y Combinator announced its inaugural debate on whether governments should slow AI development, scheduled for Monday, 11/16 at 6 p.m.; Jacob Coxon, formerly of Anthropic and OpenAI, will argue for intervention, while GMU’s Robin Hanson will argue against it.
Scott Kupor predicts SI will accelerate new-company formation, saying VC-backed company data already show the effect and the larger expected wave is in non-VC-funded small businesses.
World Labs’ Atlas was ranked #1 for camera movement in a comparison of 19 world and video models, evaluated on 600 prompts with 513K+ human judgments; the outputs, matchups, judgments, and methodology were made public on Benchmark AI. Martin Casado described Atlas as state of the art for camera control.
Jason highlighted a solo founder, describing what he is building only as “something ... scary,” and asked readers to follow him; the post gives no company or product details.
Aravind Srinivas called the benchmarks “solid” in reply to Alex Albert, who said Fable 5.2 felt like a major step above Opus 5.5 for big, open-ended work. Claude’s launch post introduces Fable 5.2 and Mythos 5.2, citing improvements in coding, reasoning, and vision.
Martin Casado criticizes what he sees as a common use of the “Bitter Lesson”: pointing to a frontier lab’s substantial spending in an area as the argument for it .
Ramez claims Anthropic is gaining exponentially less capability per billion dollars spent, with capability boosts now 8× as expensive as two years ago; he says the estimate draws on a METR report and internal models. Martin Casado says this is broadly what scaling laws would predict, while clarifying that the original laws were power laws rather than exponential; he sees scaling laws as holding, but says overcoming the curve requires substantial capital.
Martin Casado strongly endorsed @poteto’s thesis that programming-language and runtime research are entering a renaissance, software testing and verification a golden age, and that it is time to rewrite software.
On Roblox, AI-assisted game development appears to be reaching players: a16z points to a growing share of top games launched in the past 90 days, surging daily Roblox MCP users and asset uploads, and a rising share of development time spent using AI, concluding that people are making games others play. a16z contrasts gaming—and Roblox in particular—with AI app development where many products have few or no users; it cautions that the trend may still be early, with creators improving or tools getting better.
Martin Casado highlighted a technical report on using frontier models for robotics training, calling it a compelling demonstration of the direction the field is heading; he linked to the report.