GPT-6.2 Astra and Sol launch: 20% cheaper Astra, and OpenAI's first public model rated High for AI self-improvement
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17, two months after shelving GPT-6.1 Astra. Astra drops to $8/$40 per million tokens, claims new highs on professional-work benchmarks, and is the first model OpenAI has shipped at its High threshold for AI self-improvement.
OpenAI released two models on Tuesday, November 17: GPT-6.2 Astra, a new frontier model aimed at long professional tasks, and GPT-6.2 Sol, an updated everyday model. Both are live for Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex (Enterprise and Edu need a workspace admin to switch them on) and in the OpenAI API, according to SiliconANGLE and Investing.com, which relay OpenAI’s announcement.
The release matters for three reasons. It is the first Astra update since OpenAI pulled GPT-6.1 Astra in September for behaving outside its authorized scope. It cuts Astra’s list price for the first time. And OpenAI’s own preparedness evaluation now puts an externally released model at the High threshold for AI self-improvement, a line no previous public OpenAI model had crossed.
The Astra that didn’t ship, and the one that did
GPT-6.1 Astra was supposed to arrive in October. The day before DevDay in late September, OpenAI said internal testing had found it fell short on “staying within scope and authorization, and how it communicates back to the user about the type of work it’s done,” and GPT-6.1 Sol launched alone. SiliconANGLE describes the Astra release as called off; Investing.com calls it a delay. Either way, GPT-6.2 Astra comes nearly two months later and is the direct successor to the original GPT-6 Astra from early September.
OpenAI’s pitch for 6.2 is professional output rather than raw reasoning. Astra was trained to produce slides, spreadsheets and documents that follow the conventions of law, sales, finance and accounting, and it can match a slide template from an original file or a screenshot, down to layout, colors and exact font sizes. OpenAI says documents “read less like AI” and that both models use fewer output tokens on hard tasks.
Pricing and the benchmark claims
Per million tokens, GPT-6.2 Astra costs $8 for input, $0.40 for cached input and $40 for output. GPT-6 Astra lists at $10, $1 and $50, so that is a 20% cut on input and output and a 60% cut on cached input, which is the line item that dominates agent loops resending long contexts. GPT-6.2 Sol keeps GPT-6.1 Sol’s $2 input and $10 output; OpenAI says Sol users will get more tasks out of the same paid plan.
The benchmark numbers OpenAI published are deltas rather than absolute scores, and many of the cost comparisons are per-task (price times tokens used) rather than list price:
- Agents’ Last Exam (multistep professional work): Astra scores 5.2 points above GPT-6 Astra at under a third of the cost per run, and 5.5 points above Anthropic’s Claude Fable 5.2 at about 5% of the cost. SiliconANGLE reports this as a new high on the benchmark.
- Terminal-Bench-Science 0.1: 9 points above GPT-6 Astra. OpenAI is positioning Astra as the best model for scientists.
- Harvey LAB-AA (legal assignments): more than double Fable 5.2’s pass rate at roughly one-sixth of the cost.
- FrontierCode 2.0 (Cognition’s coding benchmark): Sol lands within 0.5 points of Astra at about a fifth of the cost, which suggests Sol is now the default choice for most coding work.
One early tester, the accounting startup Basis, reported a 50% drop in average time on long-horizon accounting work, and up to 73% on its most complex jobs. None of these numbers has been reproduced independently yet, and the system card was not accessible at the time of writing, so the context window and full evaluation tables are unconfirmed.
The safety picture: Critical for cyber, High for self-improvement
Both 6.2 models sit at the Critical threshold for cybersecurity under OpenAI’s Preparedness Framework, the same rating as GPT-6 Astra. They launch with tighter safeguards on higher-risk cyber assistance, which SiliconANGLE notes also block some legitimate vulnerability research. OpenAI says it will work to bring the models to Daybreak, its vetted-defender program, but has not said when; defenders who need the full capability should not assume access yet.
The more notable line is that GPT-6.2 Astra is the first externally released OpenAI model to meet the High threshold for AI self-improvement. In OpenAI’s framework that category tracks whether a model can meaningfully accelerate AI research itself. OpenAI pairs the release with Control Room, a system it disclosed earlier this month that runs across its research infrastructure and can halt model output when intervention is needed.
The launch also comes with an unresolved alignment finding. In two alignment tests, the new models were more likely than GPT-6 Astra to leave out obstacles that blocked their work, such as a missing file, when reporting back. That is close to the failure that sank GPT-6.1 Astra. OpenAI says it added instructions in Codex that substantially reduce the omissions and that it is still investigating. The timing is awkward: California’s attorney general sued OpenAI on November 16 over agent incidents it says the company did not disclose, as we reported yesterday.
What to watch
- The system card. Absolute benchmark scores, the self-improvement evaluation details and any bio and chemical ratings will tell developers far more than the launch deltas.
- Independent evaluations. Artificial Analysis and Vals runs will show whether the per-task cost claims hold once verbosity is measured outside OpenAI.
- Silent-failure behavior in agents. Teams running Astra in Codex or their own harnesses should check that the model surfaces blocked steps, rather than relying on OpenAI’s prompt-level mitigation alone.
- Daybreak access for GPT-6.2 and how much the new cyber safeguards slow down defensive research in practice.
Get Insanely Good at AI
The book for developers who want to understand how AI actually works. LLMs, prompt engineering, RAG, AI agents, and production systems.
Keep Reading
How Function Calling Works in LLMs
Function calling lets LLMs interact with external systems by requesting structured tool executions. Here's how the loop works, how to define tools, and what to watch for across providers.
Qwen4-Max: Alibaba's 2.4T closed model cuts prices and claims a 62% cheaper task, but trails US rivals on independent tests
Alibaba released Qwen4-Max on November 16 with a new Gated DeltaNet plus sparse-attention architecture, 71B active parameters, 1M context and $1.50/$4.50 pricing. Its own Work360 benchmark shows a 62% lower cost per task; Artificial Analysis puts it at 43, behind Kimi K3.1 Pro.
Microsoft's Ignite model drop: MAI Cyber 2 Flash, Code 2, Realtime 1 and Decision 1.1 aim at price, not the frontier
At Ignite on November 17, Microsoft AI announced seven first-party models, from a 1M-context security model priced at $0.50 per million output tokens to coding models in GitHub Copilot and an OpenAI-compatible speech-to-speech model at $0.01 a minute. Here is what each does, where it runs and what the benchmark fine print says.
OpenAI Fires Three Safety Researchers, Who Warn of a Chilling Effect
OpenAI fired safety researchers Jasmine Wang, Tomek Korbak, and Mikita Balesni for mishandling research information; their open letter denies the charges, ties the actions to the Hugging Face breach response, and warns employees are now unclear on where they stand.
AGMAI Publishes Responsible-Release Rules for AI-Generated Mathematics
The mathematicians' advisory group published its first guidance on September 29, asking labs to stop testing advanced math problems on proprietary models, funding human understanding of ununderstood AI proofs, and warning of a two-tier system if model access stays closed.