OpenAI has introduced GPT-6.2 Astra and Sol, two days before Anthropic’s expected first day of stock market trading. In Artificial Analysis tests, Sol matches the performance of September’s Astra at roughly 15% of the cost per task. The new Astra scored 62 on the Intelligence Index, behind Fable 5.2 at 65.
The models are rolling out in ChatGPT Work and Codex to Plus, Pro, Enterprise and Edu customers, while developers can access them through the API as gpt-6.2-astra and gpt-6.2-sol. Both have a one-million-token context window and a 128,000-token output limit.
Per million tokens, Astra costs $8 for input, $0.40 for cached input and $40 for output. Sol retains its predecessor’s prices of $2, $0.10 and $10 respectively. Astra for Law, an Enterprise configuration, adds a legal search index: Artificial Analysis measured it at 20.8% on Harvey LAB-AA, compared with 15.0% for the base model.
Benchmarks and safety
OpenAI says Astra, at its xhigh setting, scored 5.2 points above GPT-6 Astra at its maximum setting on Agents’ Last Exam 1.1, while using less than a third of the cost per task. On the same test, the company says the new model led Fable 5.2 by 5.5 points at about 5% of its cost. On GDPval v2, it used 22% fewer output tokens than its predecessor.
On Mercor’s OpenWorldEngineer, Astra scored 67.2%, Fable 5.2 scored 54.6%, and GPT-6 Astra scored 44.2%. Astra ranked first and Sol third on Vals’ multi-turn Vibe Code Bench. In AA-3D, Astra led with 1,318 points, ahead of Fable’s 1,271 and Sol’s 1,259; on Artificial Analysis’s robotics preview, their respective scores were 79, 75 and 78. On Perplexity’s WANDR research-agent test, Astra scored 0.783 at approximately $12.54 per task, while Sol scored 0.746 at about $2.65.
Epoch verified seven solutions out of 68 FrontierMath Erdős problems in Lean, up from a previous best of three. OpenAI rated Astra High for AI Self-Improvement, the first such rating for one of its externally released models. In a separate measure, Vals’ RSI-Index rose from 27.07% to 29.01%; its human reference is 50%.
Mixed early reception
The launch also drew strongly negative reactions. After weeks of testing, Matt Shumer called the models essentially unusable and said his Opus and Fable agents rejected much of the work produced by Astra subagents. Pietro Schirano attributed many of his initial problems to the agent harness. Arena’s Peter Gostev highlighted Astra’s SVG peacock in a visual coding comparison.
Among the demonstrations, Gabriel Chua built a walkable Singapore using more than 100 Codex threads, while Simon Willison ran the same three.js mech-building prompt through several models. The source’s authors said some criticisms were valid and that artifacts of reinforcement learning were noticeable, but they continued to use the models to automate their work.
Microsoft, Google and agent infrastructure
Microsoft introduced its own coding models: MAI Code 2 for complex, multi-step tasks and Code 2 Flash for faster edits. Both are available in the GitHub Copilot model picker, Copilot Fusion and Foundry. The company also announced a multimodal upgrade to its Decision model. It says Copilot Fusion’s multi-model harness scored 71.2% on Terminal-Bench 4.1 at 88% lower cost than Claude Fable 5.2 in that comparison. Artificial Analysis ranked MAI Image 2.7 first on its image-editing benchmark, at $34.82 per 1,000 images.
Google says Hydrogen powers free Gemini Tasks in six countries, including tasks requiring a phone call or text message. Google DeepMind also showed an Antigravity robotics plugin using MuJoCo simulation.
A small Hugging Face adapter trained inside pi increased Qwen3.8-27B’s solves on the 51 hardest SWE-bench Pro tasks from 15 to 32, with 55% fewer tool calls per task. Baseten found up to a 4.4-fold difference in cost per passing task across a benchmark of 22 harnesses and five open models. Google’s SIM research suggests harnesses may also specialize to a model: a harness evolved on Opus scored 68.41 on Gemini, compared with 73.89 for the starting harness.
Cognition launched persistent Devin Bots. Perplexity says its mixed-model teams beat four instances of the lead model on Terminal-Bench and WANDR. Replit found diminishing returns when increasing agent counts from 16 to 32 on NumPy tasks. In a 1,000-agent Hugging Face run, only 40–60 agents called the model at any moment; they read roughly 130 tokens for every token written, while merges and tests also created bottlenecks. In a SemiAnalysis sample, 1.8% of requests not served from cache accounted for roughly 15% of spending.
Open models and acceleration
Marin released 57 intermediate Mochi checkpoints from its 535B-A23B training run. A K2 Horizon report compares four model sizes trained on the same 22 trillion tokens and describes a decoding adapter that the team says achieves a 2.2-fold speedup. After converting several models to its subquadratic attention mechanism, Zyphra reported 9–60-fold speedups at 10 million tokens; Qwen lost fewer than two points on RULER at one million tokens.
The vLLM project says a sliding-window draft model for Kimi K3.1 Pro increased accepted sequence length in speculative decoding by up to 30%. NVIDIA also has a four-billion-parameter draft model for Qwen3.8-27B.
SenseNova-Loop-9B runs 24 of its 32 layers a second time. Its authors report that its GAIA score rose from 37.86 in the preview to 69.90 in the release; local users asked how much memory remains for the KV cache at long context. Maxwell V1 Flash, a 35-billion-parameter model with three billion active parameters, claims a 6.7-fold attention speedup over dense GQA at 256,000 tokens, but its sparse attention still retains the full-history KV cache. OVHcloud previewed Olala mini 7A1B, with one billion active parameters and a 32,000-token context window; weights and reproducible benchmarks are still awaited.
An eye-catching throughput comparison in Intel’s early Crescent Island brief uses FP64-heavy workloads. In a separate LLM comparison, Intel claims 1.12 times the throughput and 29% lower latency than the RTX PRO 6000 Blackwell Server Edition. A developer also reported running Gemma 3 270M through a PostScript print job.
Companies and applications
In its first earnings report after its CEO transition, NVIDIA reported quarterly revenue of $115.8 billion under interim Chief Executive Officer Colette Kress, up 103% year over year. The company said Vera Rubin NVL72 had shipped, while its board said it did not intend to reinstate Jensen Huang.
Isomorphic Labs raised $5 billion in a Series C co-led by Greenoaks and Sequoia to scale its drug-design platform and therapeutic development pipeline. Genesis AI raised $500 million in a Series A that valued the robotics company at $3.5 billion after the investment. It plans to use the funds to deploy its Eno robot and expand training and data collection.
The miHoYo Gamer video-game agent samples the screen five times a second and emits keyboard and mouse actions at 30 Hz. In its authors’ evaluation, the reinforcement-learning version solved about half of the held-out Honkai puzzle tasks. Another developer says the fictional Fock Ball league simulates matches in a physics engine and uses Claude for work such as commentary and the newsroom.
In a discussion about ICLR peer reviews, reviewers reported that program chairs had flagged reviews likely to have been generated by AI. One question raised was how a detector could distinguish editing assistance from the generation of an entire review.



