OpenAI announced GPT-6.2 Astra and GPT-6.2 Sol on November 17, 2026. The next day, OpenAI Developers announced a public beta of background mode, with two- or eight-hour windows and savings of up to 40% per token versus Standard for qualifying workflows. The launch announcement and beta announcement describe separate releases.
Compare the models by the work you need done, and keep API rates, background-mode terms and benchmark claims in their own lanes.
Bloomberg reported that OpenAI had decided in late September not to release GPT-6.1 Astra, citing concerns about scope, authorization and how the model communicated its work. That earlier decision was reported by Bloomberg in an account republished by Yahoo Finance España. NeoTeo covered the earlier GPT-6 Astra release.
What the November 18 API beta adds
OpenAI Developers announced a public beta for background mode on November 18, with two- or eight-hour windows and up to 40% savings per token compared with Standard for qualifying workflows. The announcement describes beta terms, not a change to the published standard token rates.
Where Astra and Sol are available
OpenAI lists GPT-6.2 Astra and Sol for Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex, as well as through the API under the model names gpt-6.2-astra and gpt-6.2-sol. The rollout does not add them to regular Chat. Access depends on plan eligibility, workspace permissions and rollout progress, according to OpenAI’s launch details and community clarification.
OpenAI also specifies Ultrafast access for Pro 500 subscribers with either model; Pro 200 access begins with Sol. Eligible Enterprise and Edu workspaces receive that access after an administrator enables it. These conditions matter if a model is missing from an account: the plan and workspace settings affect access.
Published API rates for Astra and Sol
OpenAI’s published API rates are in U.S. dollars per million tokens. Standard input, cached input and output each have a separate rate:
| Model | Standard input | Cached input | Output |
| GPT-6.2 Astra | $8.00 | $0.40 | $40.00 |
| GPT-6.2 Sol | $2.00 | $0.10 | $10.00 |
Each cached-input rate is one-twentieth of that model’s standard input rate. The background-mode beta’s potential per-token savings are a separate offer for qualifying workflows; they do not replace these listed rate categories. OpenAI says Astra’s standard input and output rates are 20% lower than GPT-6 Astra’s, and its cached-input rate is 60% lower. OpenAI lists the rates and comparison.
What OpenAI reports about performance and safety
OpenAI positions Astra for difficult professional work and Sol as an efficient model for everyday tasks. Its launch page reports Astra scored 5.2 points above GPT-6 Astra on Agents’ Last Exam at less than one-third the cost, and 5.5 points above Claude Fable 5.2 at 5% of its cost. The benchmark evaluates long-horizon professional workflows; OpenAI notes that Fable 5.2’s leaderboard excludes nine tasks that used a fallback model.
OpenAI also reports Astra scored 9 percentage points above GPT-6 Astra on Terminal-Bench-Science 0.1 at less than one-third the cost. On GDPval v2, OpenAI reports a 6.2-percentage-point improvement over GPT-6 Astra using 22% fewer output tokens; it reports a 5.3-point improvement for Sol over GPT-6.1 Sol. These are OpenAI’s benchmark results, each tied to its stated test and comparison.
OpenAI classifies both models at its Critical threshold for cybersecurity and High threshold for biological and chemical capabilities. It places Astra at High for AI Self-improvement as well. OpenAI says it applied tighter safeguards, which can restrict some legitimate vulnerability research and security testing. These are the company’s assessments under its Preparedness Framework.
What Every’s early review found
Every’s Dan Shipper and his team criticized Astra’s writing as overly compressed, said it sometimes stopped early on end-to-end coding tasks, and found limited variation in default design outputs. The review also showed a 3D town made using projected 2D images, with surface details disappearing from some angles. Every disclosed that its team had early access and free token usage from AI labs.
Those observations concern different tasks from OpenAI’s benchmarks. OpenAI employee Tibo said some early-access videos may have reflected testing before changes made in the final three to four days before launch; he also acknowledged remaining work on usability and taste. In the OpenAI developer community, one participant reported that Sol found more issues across a coding project than GPT-6.1 Sol and improved their 3D modeling results, while still requiring iterations.