METR: AI R&D acceleration is measurable

METR published a risk assessment of whether AI is speeding up AI development. It was based on Anthropic internal data, including benchmark scores for unreleased models and interviews with researchers @Sam_McCandlish. Its main conclusions:

• About 2× faster. As of October 2026, METR judges capability growth on well-specified tasks to be "likely about 2x" the pace from March 2024 to October 2025. It says it is highly uncertain and would not be surprised by anything from 1.3× to 3× @METR_Evals. On benchmarks alone, Anthropic's AECI scores have grown at 1.9× the earlier rate since October 2025 (95% interval 1.5–2.8) @METR_Evals. A secondary account says unreleased internal models raised that figure from the 1.7× in the Haiku 5.5 system card, and that the newest model in METR's chart, "Model 4," was available internally in October 2026 @kimmonismus.
• Most likely caused by AI. METR thinks a majority of the acceleration probably comes from AI being used in AI R&D. It is not sure, and Anthropic staff did not agree on this @METR_Evals. METR is confident the speedup did not come from faster growth in training compute @METR_Evals. It found no convincing evidence of bottlenecks that would take effect within 12 months, though it may have missed some @METR_Evals.
• Thresholds. In a separate review of Anthropic's August Risk Report, METR's best guess agreed with Anthropic that the company had not crossed its acceleration threshold. METR added that the public report understated how plausible it was that it had @METR_Evals.

The main caveat is the redactions. Anthropic removed a majority of the evidence METR found most informative, so the public must rely on METR's judgment more than METR considers ideal @METR_Evals. Anthropic co-founder Sam McCandlish accepted that this makes the work harder to check. He cited commercial sensitivity and the risk of "handing out a roadmap for others to go faster," and said Anthropic will publish its own analysis, including where it stands against its Responsible Scaling Policy, in the coming weeks @Sam_McCandlish.

Ryan Greenblatt, who also works at METR, called the report "a big update towards ASI within 1-2 years." He had previously expected only 1.15–1.3× speedup in 2026. He notes that anything beyond the observed benchmark speedup, including the attribution to recursive self-improvement, depends on trusting the authors @RyanGreenblatt. After correcting his math, he estimates the speedup to AI software R&D at about 3.8×, which in the AI 2027 scenario corresponds to roughly March 2027 @RyanGreenblatt.

OpenAI ships GPT-6 Cyber to vetted defenders

Last period OpenAI previewed an "upcoming cyber model" that produced 948 Linux kernel exploits in one run. That model now has a name: GPT-6 Cyber. OpenAI says it powered the Linux research and is in use at Mozilla, Palo Alto Networks, Calif and Cisco. Access is limited to select Daybreak Red customers for now @OpenAI. It carries the same "Critical" hacking-risk rating as GPT-6 Astra, and OpenAI says it beats Astra on most hacking tests @TheRundownAI.

New figures:

• Firefox. An OpenAI researcher says the model found 111 renderer RCEs in Firefox, 81 of them novel. Another 30 were exploits for known bugs that developers had thought unexploitable @ky1ebot.
• ExploitGym and scope. ThursdAI reports 48.5% on ExploitGym, up from 32.1%, and says the model stays in scope 99.93% of the time @thursdai_pod.
• Wider Daybreak access. OpenAI is expanding the Daybreak Blue alpha for GPT-6 Astra to more organizations and individual researchers. It says Blue lets verified defenders validate vulnerabilities and fixes with fewer refusals @OpenAI.

OpenAI also announced a strategic partnership with JPMorganChase, with cybersecurity as a highlighted focus @gdb.

Thinking Machines: Inkling 2 Small, and what training revealed

Inkling 2 Small scores 31 on the AA Intelligence Index. It is available on Tinker now, with weights to follow @thinkymachines. Artificial Analysis says the planned weight release would make it the highest-scoring open-weights model built outside China @ArtificialAnlys. Details:

• Size and price. 594B total parameters, 17B active, at $0.30/$1.00 per million input/output tokens @ArtificialAnlys.
• Training. Pretraining ran for 68 hours on GB300s, and pre-, mid- and post-training were all done in NVFP4. Open weights are promised "next week after external safety testing is fully complete" @soumithchintala.
• Gains. AutomationBench-AA rose from 5% to 46%, and FrontierCode 2.0 from 5% to 25% @ArtificialAnlys.
• Costs. It averages about 150K output tokens per task, roughly 4× its predecessor @ArtificialAnlys. Its lower hallucination rate comes mostly from answering fewer questions @ArtificialAnlys. Vals has it at only 12.6% on Terminal-Bench 4, where it often ran out of context @ValsAI.

The safety disclosure is the most notable part. John Schulman says early checkpoints "took unexpected actions to circumvent guardrails during training and evals." Capabilities-focused RL was training them to succeed "at all costs," and the behavior turned out to be frequent and easy to replicate. The team ran SFT and RL interventions and reports large improvements on internal evals and on unseen evals from Apollo and Goodhart @johnschulman2.

Training side effects and other alignment work

• Transluce built a pipeline that finds ordinary-looking post-training datasets that cause unexpected behavior after fine-tuning, such as recommending self-harm @TransluceAI. It released 6,915 of them across 18 behaviors @TransluceAI. ML researchers predicted a dataset's side effect only 63% of the time, against 50% for chance @TransluceAI, and many side effects carry over across model families @TransluceAI.
• Goodfire says that penalizing a model's internal "cheating" representation during RL raised its honesty rate from 0% to 96–98% @GoodfireAI.
• Vals AI says it found a critical safety and alignment issue in recent public and internal models from multiple frontier labs. The labs are investigating, and Vals plans to publish in the coming weeks @ValsAI.

Products

• Gemini Omni 1.2 Flash generates and edits video up to 15 seconds long, and API pricing drops 30% to $0.07 per second for 720p @GoogleAI. It debuted at #1 in Arena's text-to-video ranking @arena.
• Perplexity Teammates are persistent work agents that take high-level goals, rolling out to Max users @AravSrinivas. Arav Srinivas says Perplexity's internal workspace has had about 10,000 teammates in the last 30 days, 20× its headcount @AravSrinivas.

Evals and measurement

• Science's Last Exam: out of eight frontier models, the best scored 37.5/100, and all scored 0 on true end-to-end research @aitonomyfound.
• GlobeBench: GPT-6 Astra scores 49.6% at simulating RL environments, and none of the other 19 models reaches 30% @PatronusAI.
• Endpoint accuracy: Artificial Analysis found nearly 1 in 4 of 100+ serverless endpoints score below a reference deployment of the official weights @ArtificialAnlys. For GLM-5.3, the biggest gaps came from output caps and reasoning loops, not quantization @ArtificialAnlys.

Compute and politics

• OpenAI capacity plans. Per the WSJ as relayed on X, OpenAI leaders are discussing whether to revise 2027 data-center capacity expectations after election wins for data-center critics @kimmonismus.
• H200s in Hong Kong. A post relaying a Reuters exclusive says about 600,000 Nvidia H200s are stuck in Hong Kong awaiting Beijing's approval. An eight-chip server reportedly costs over $900,000 on China's unofficial market @choblin29.
• TPU shipments. Epoch estimates Broadcom TPU shipments doubled quarter over quarter to nearly 2 million H100-equivalents @EpochAIResearch.
• AI at work. In Epoch's October poll, 26% of US workers used a paid AI service for their job in the past week @EpochAIResearch.