Guozhen AIGlobal AI field notes and model intelligence

Realtime AI News

OpenAI ships GPT-6.2 Astra and Sol together, pushing flagship quality down-market

OpenAI released GPT-6.2 Astra and GPT-6.2 Sol at the same time, pairing a flagship built for hard projects with a lighter model for high-frequency professional work that the company says comes close to flagship quality. Astra tops the AA-3D benchmark at 1318 Elo, and its API price fell 20% as both models went live in ChatGPT Work, Codex and the API.

Published

OpenAI released two models on November 18: GPT-6.2 Astra and GPT-6.2 Sol. QbitAI described the launch as a double release, with Astra aimed at hard projects and Sol handling high-frequency, everyday professional tasks — one version number, two very different cost structures.

The segmentation is the story. Astra is the flagship for the hardest work, while Sol is positioned as a lighter model that approaches flagship performance on tasks run many times a day. The choice for users is no longer simply which model is best, but whether a given step justifies the most expensive one.

On published evaluations, Astra's gains cluster around complex tasks, research and coding. It improves 5.2 percentage points over GPT-6 Astra on Agents' Last Exam while cutting per-task cost to less than a third, according to the report, and adds 9 points on Terminal-Bench-Science 0.1.

In coding, Astra scores 44.2% on FrontierCode 2.0, up 1.4 points. Sol trails it by only 0.5 points at roughly a fifth of the cost. That pairing is the most practical part of the release: most development work does not need flagship precision, but it does need a price that survives repeated calls.

3D generation is the other contested front. Astra takes first place on Artificial Analysis' AA-3D leaderboard with 1318 Elo, ahead of Claude Fable 5.2 at 1271, and Sol reaches 1259 — the result behind the article's claim that 3D work is in trouble again.

Office work gets its own section. Astra improves 6.2 points and Sol 5.3 points on GDPval v2, and Astra needs 22% fewer output tokens to finish a task. The examples include building a revenue forecast and valuation table from financial data, and filling a one-page market analysis deck with brands, price bands and product images.

Pricing is the hardest signal here. Astra's API rate drops 20% to $8 per million input tokens and $40 per million output tokens, with cached input down 60%; Sol holds at $2 and $10. Both are live in ChatGPT Work, Codex and the API for Plus and Pro subscribers.

There is a safety note as well: the report says both models upgrade vulnerability discovery, exploitation and code repair, and both reach OpenAI's highest cyber-risk tier. On the DeepBinary bug-hunting evaluation, the bounties for flaws Astra found are estimated at more than $1 million, against about $440,000 for the previous generation.

The open question is whether the flagship-plus-light split becomes permanent. If Sol holds close to Astra on real development work, model choice turns from a capability question into a cost question, and pricing and caching policy will matter more than leaderboard position in deciding what gets called every day.

Why it matters

Shipping a flagship and a near-flagship together lets OpenAI compete on cost per task rather than benchmark prestige, which shifts pressure onto rivals' pricing as much as their model quality.

OpenAIModel ReleaseAgent
Back to realtime news

Nearby Updates

All

11/18, 08:16

Nvidia says record revenue run won't stop, even amid leadership turmoil

Nvidia said its string of record-breaking revenue is not about to end, even as the company navigates the fallout from founder and CEO Jensen Huang's departure, TechCrunch reported. The report attaches a caveat to the optimism: the projection is framed as holding for now.

11/18, 08:00

Reson8 open-sources Rumble-2-Flash, a 0.5B speech model for 40+ languages

Reson8 released Rumble-2-Flash on November 18 under Apache 2.0: a 0.5B-parameter speech recognition model covering more than 40 languages that runs natively in Transformers and vLLM. It posts a 6.96% average error rate on the Open ASR Leaderboard's multilingual track, beating Meta's 15x larger omniASR-LLM 7B and matching a 30B Nemotron model on a 302-hour call-centre benchmark.

11/18, 08:00

OpenAI and Groupe Figaro partner to bring Le Figaro journalism into ChatGPT

OpenAI and Groupe Figaro announced a strategic partnership that makes Le Figaro's current reporting and archive content discoverable inside ChatGPT through summaries and excerpts with links to the original stories. Both sides said the publisher keeps full editorial independence and control over its journalism.

11/18, 08:00

OpenAI details model-driven chip design: 13% area savings on Jalapeño, 14% PPA gain on AMD GPUs

OpenAI described how its models cut standard-cell area by roughly 13% in Jalapeño's warp engine and 10% in its matrix unit, and delivered up to 14% PPA improvement over heavily optimized baselines on AMD's flagship GPU IP. It also reported GPT-6.2 Astra leading evaluations on two open-source RISC-V cores and a new partnership with Synopsys to train GPT-Synopsys on EDA tools.