GPT-6.2 Astra and Sol: where OpenAI gained, where it lost, and what it costs
By AgentRiot Editorial
OpenAI’s new flagship finishes near the top of the independent intelligence index at a fraction of Claude Fable 5.2’s measured cost. Sol makes the stronger case for routine work, but its Codex regression is a reason to test before switching.

OpenAI’s best result on Agents’ Last Exam, a test of long-running professional work, came from turning GPT-6.2 Astra down. At xhigh reasoning, it scored 78.33% at $2.89 per task. At max, it scored 77.61% at $4.47. The difference in score is small; the bill is not. The old GPT-6 Astra reached 73.14% at $8.99. That is a better introduction to the pair than the games in @ChatGPT’s November 17 announcement (opens in new tab): with these models, the setting matters almost as much as the name.
OpenAI launched GPT-6.2 Astra and Sol (opens in new tab) for ChatGPT Work, Codex and the API. Astra’s standard input and output prices fell 20% from GPT-6 Astra. Sol kept its predecessor’s rates. The most useful question is whether those rates buy more finished work, not whether one model wins the most rows.
The independent scorecard: Fable still leads
Artificial Analysis’s launch assessment (opens in new tab) places Astra second among model families on its Intelligence Index v5.0.1, a nine-evaluation composite covering areas such as coding, science and knowledge work. The top Fable 5.2, Astra, Carbon and Opus configurations are separated by 5.5 index points, but their weighted test costs are much farther apart. Here are the operator’s published Astra (opens in new tab) and Sol (opens in new tab) comparison figures, including the outgoing Astra (opens in new tab) and Sol (opens in new tab) baselines. Fable and Opus are the operator’s default-fallback configurations.
| Model and tested effort | Index, points | Weighted API cost per task |
|---|---|---|
| Claude Fable 5.2, Max | 64.6 | $26.54 |
| GPT-6.2 Astra, Max | 62.1 | $4.00 |
| Gemini 4 Carbon, High | 60.0 | $5.48 |
| Claude Opus 5.5, Max | 59.1 | $10.04 |
| GPT-6.2 Sol, Max | 57.3 | $1.01 |
| GPT-6 Astra, Max | 56.9 | $6.24 |
| GPT-6.1 Sol, Max | 54.6 | $1.27 |
Astra costs about 15% as much as Fable on that particular test mix. Sol matches old Astra when the index is rounded to whole points and costs about 16% as much. Neither ratio predicts an individual team’s bill. The index weights several benchmarks together, and models can use very different amounts of input, cache and reasoning output to solve the same task.
The operator’s component results (opens in new tab) explain why Fable still leads overall. Astra beat it on Terminal-Bench-Science (78.1% vs. 73.3%) and Terminal-Bench 4.0 (64.1% vs. 59.6%), though Carbon edged Astra on the latter (65.2%). Fable beat Astra on HLE-Diamond (80.2% vs. 72.2%), hard CritPt v2 physics tasks (76.5% vs. 71.5%), and AA-Briefcase (1,785 vs. 1,644 Elo). Sol registered 67.6% on Terminal-Bench-Science, 60.1% on Terminal-Bench 4.0, 64.0% on HLE-Diamond and 62.5% on hard CritPt v2. The two GPT-6.2 models were nearly tied on the strict GDP-PDF measure, at 29.8% for Astra and 30.0% for Sol. That measure is separate from OpenAI’s internal GDPval v2.
The factual-reliability results deserve their own line. On AA-Omniscience v1.1 (opens in new tab), Astra answered 68.6% correctly and Sol 61.6%, compared with Fable’s 79.1% and Carbon’s 58.3%. Carbon was much less inclined to guess: its hallucination rate on non-correct responses was 11.4%, versus 59.1% for Astra and 55.4% for Sol. Those are not error rates across all answers. The metric distinguishes an outright wrong answer from a partial answer or an admission of uncertainty.
OpenAI’s launch tests: gains in documents and science, losses in coding and CAD
The following figures come from the charts in OpenAI’s launch report (opens in new tab), not an independent reproduction. Each cell shows the best score plotted for that model and its associated cost per task, or output-token count where the chart uses tokens. A dash means the model was not plotted. For ties, the lower-cost point is shown.
| Benchmark | GPT-6.2 Astra | GPT-6.2 Sol | GPT-6 Astra | GPT-6.1 Sol | Claude Fable 5.2 | Gemini 4 Carbon |
|---|---|---|---|---|---|---|
| Agents’ Last Exam v1.1 | 78.33% / $2.89 | 73.04% / $1.32 | 73.14% / $8.99 | 72.45% / $2.02 | 72.79% / $60.05 | 77.41% / $4.60 |
| Terminal-Bench-Science 0.1 | 77.1% / $7.15 | 68.0% / $2.21 | 68.1% / $23.80 | 57.0% / $5.47 | 73.3% / $45.21 | - |
| FrontierCode 2.0 | 44.2% / $5.97 | 43.8% / $1.29 | 42.8% / $6.68 | 39.0% / $1.03 | 49.11% / $14.19 | 41.1% / $5.16 |
| LifeVal | 86.1% / $4.23 | 83.8% / $1.27 | 79.8% / $9.04 | 79.5% / $1.27 | - | - |
| Harvey LAB-AA v1.1, strict pass | 15.0% / $6.35 | 11.1% / $1.92 | 8.6% / $13.67 | 6.9% / $2.48 | 7.2% / $39.84 | - |
| GDPval v2, internal | 50.4% / 35,812 tokens | 46.9% / 33,428 tokens | 44.2% / 45,732 tokens | 41.5% / 40,428 tokens | 44.8% / 222,344 tokens | 48.8% / 108,729 tokens |
| Vals Science Index v1 | 68.4% / $4.17 | 58.5% / $0.71 | 65.1% / $8.03 | 59.6% / $1.36 | 63.74% / $8.79 | - |
| BenchCAD 2.1 Core | 82.5% / $5.43 | 73.1% / $1.60 | 62.6% / $9.71 | 52.6% / $1.61 | 83.5% / $50.77 | - |
| AutomationBench 1.0.6 | 52.4% / $0.83 | 41.2% / $0.21 | 41.4% / $1.73 | 36.1% / $0.30 | 51.45% / $4.64 | 53.9% / $2.12 |
On professional work, Astra’s best Agents’ Last Exam score was 5.19 percentage points above old Astra’s at 32% of the cost. It beat Fable by 5.54 points at roughly 5% of the cost, but OpenAI notes that the Fable leaderboard result excludes nine tasks run with a fallback model. Carbon was within one point of Astra. On internal GDPval v2, Astra gained 6.2 points while using about 22% fewer output tokens than its predecessor. The displayed Sol results differ by 5.4 points (46.9% vs. 41.5%); the launch text reports 5.3 points. On the legal benchmark, a deliverable has to meet every grading criterion without material hallucinations to pass. A 15.0% strict pass rate leads this comparison; it still leaves most deliverables short of the bar.
On science, Astra gained 9.0 points over old Astra on Terminal-Bench-Science at 30% of the cost. That’s a test of research workflows in a terminal, not a measure of discovery in a lab. Fable’s 73.3% row comes from Artificial Analysis’s mini-swe-agent run; OpenAI’s footnote (opens in new tab) gives Anthropic’s separate 74.6% result in Claude Code. On the Vals Science Index (opens in new tab), Sol’s best launch-chart result, at xhigh, was cheaper than GPT-6.1 Sol’s max run but scored 1.1 points lower. The upgrade did not help every scientific workload.
On software and automation, Sol came close to Astra on the FrontierCode chart at 22% of the cost, but Fable led both, and Opus 5.5 beat both at 45.5% and $12.82. Cognition’s FrontierCode methodology (opens in new tab) grades whether maintainers would accept a change in a real codebase; a weighted rubric score is not the percentage of pull requests merged. The native-agent result gives a separate warning: Artificial Analysis reports that Codex at max effort was flat with new Astra and down three Coding Agent Index points with new Sol versus their predecessors. Sol lost 14 points on SWE-Atlas-QnA-AA, where it was more likely to leave out key facts. Separately, Fable narrowly beat Astra on CAD, and Carbon narrowly beat Astra on business automation. Astra was cheaper in both cases.
The demo that best explains the work pitch is a retailer’s earnings deck. OpenAI gave both Astras the same Q1 deck, Q2 earnings release and 10-Q. Both spotted tariff refunds as the main driver of the earnings gain. The newer model put 5.3% sales growth in the headline, separated online and in-store growth, and explicitly called out budgeting without another refund. That is a vendor-selected example, not a pass-rate estimate. For a harder check on repeatability, the AA-AnalystAgent model records (opens in new tab) show 62.5% for Astra and 53.75% for Sol on the metric that requires a correct answer in all five attempts. In Artificial Analysis’s launch results (opens in new tab), Astra also led the human-rated AA-3D comparison at 1,318 Elo, ahead of Fable at 1,271 and Sol at 1,259. On the separate AA-Robotics v0.1 research preview, Astra, Sol and Fable scored 79, 78 and 75 respectively. Those results don’t certify a CAD assembly or a deployed robot.
The ARC-AGI headline depends on the harness
ARC Prize’s Astra results (opens in new tab) and Sol results (opens in new tab) report separate semi-private scores. ARC-AGI-3’s Standard harness lets the model carry forward notes it chooses to keep. The Provider Adapter also preserves opaque reasoning state between calls and compacts long conversations. The adapter result therefore describes a specific model-and-harness combination.
| ARC Prize semi-private result | GPT-6.2 Astra | GPT-6.2 Sol |
|---|---|---|
| ARC-AGI-1, Max | 98.5% | 97.5% |
| ARC-AGI-2, Max | 95.8% | 85.8% |
| ARC-AGI-3, Standard, Max | 70.3% | 54.9% |
| ARC-AGI-3, Provider Adapter, best | 100.0% at Xhigh | 98.2% at Max |
Astra scored 99.8% with the adapter at Max, so 100% is an Xhigh result, not a higher-effort improvement. ARC Prize also says Sol coverage is partial: two v1/v2 tasks and eight v3 games remain incomplete across the published evaluations. Its reported scores use the full dataset denominators, while its recorded costs cover only usage that was run.
API pricing: the context threshold matters
OpenAI’s standard API rates (opens in new tab) are in dollars per million tokens. Short context means up to 272,000 input tokens per request. Once a request exceeds that threshold, the higher rate applies to the whole request, not just to the tokens above the line.
| Standard API rates | Uncached input | Output | Cache read | 30-minute cache write | 12-hour cache write |
|---|---|---|---|---|---|
| GPT-6.2 Astra, short | $8 | $40 | $0.40 | $10 | $20 |
| GPT-6.2 Astra, long | $16 | $60 | $0.80 | $20 | $40 |
| GPT-6.2 Sol, short | $2 | $10 | $0.10 | $2.50 | $5 |
| GPT-6.2 Sol, long | $4 | $15 | $0.20 | $5 | $10 |
Astra’s cache-read rate fell 60%, from $1.00 to $0.40 at the short-context rate. Sol’s rate card is unchanged from GPT-6.1 Sol. Batch is half the Standard rate and Fast is twice it. Ultrafast is three times Standard through February 28, 2027, rising to six times Standard on March 1, Pacific time. At short context, that takes Astra’s promotional $24 input/$120 output to $48/$240, and Sol’s $6/$30 to $12/$60. The model specs for Astra (opens in new tab) and Sol (opens in new tab) list a 1.05 million-token context window, up to 922,000 input tokens and up to 128,000 output tokens. Both accept text and images, return text, and reject the none reasoning setting.
For a fixed workload of 1 million uncached input tokens and 250,000 billed output tokens, spread across individually short requests and including reasoning in the output bill: Sol costs $4.50, Astra $18.00, and old Astra $22.50. At Anthropic’s published standard rates (opens in new tab), the same workload costs $18.00 on Fable 5.2 and $9.00 on Opus 5.5. Gemini 4 Carbon (opens in new tab) costs $4.50 through November 30 and $9.00 from December 1. This calculation excludes retries, tool charges, cache creation and reads, and expedited tiers. Astra and Fable have identical base input and output tariffs in this example; their 6.6-fold cost gap in the independent index reflects how the evaluated configurations used tokens, not a sixfold list-price discount.
Access, safeguards and what to try first
The launch announcement (opens in new tab) and Codex changelog (opens in new tab) name Plus, Pro, Enterprise and Edu in ChatGPT Work and Codex. The Help Center (opens in new tab) also lists Business, Healthcare and FedRAMP as the rollout reaches eligible workspaces, and explicitly excludes regular Chat from this rollout. Eligible Enterprise and Edu workspaces need admin enablement. Pro $500 gets Ultrafast for both models; Pro $200 gets it for Sol. Through the API, the IDs are gpt-6.2-astra and gpt-6.2-sol, with separate metered Ultrafast access. Subscription usage is not priced by the API example above.
OpenAI’s system card (opens in new tab) rates both models Critical for cybersecurity and High for biological and chemical capabilities under its own framework. Astra is its first externally released model designated High for AI self-improvement. Those designations concern assessed capabilities, not typical behavior in public deployments. In OpenAI’s defensive tests, Astra scored 65.7% and Sol 55.6% on secure code generation (SecureSWE); they scored 79.7% and 75.6% on PatchBench. The alignment results are mixed: improper actions in simulated workplace tasks fell to 0.64% for Astra and 0.71% for Sol, but in a separate test the models were more likely than their predecessors to omit a blocker. The measured omission-related rate fell from 22.28% to 8.04% for Astra and from 24.08% to 8.88% for Sol with a Codex developer prompt. When a workflow depends on a source file, make the agent report which files it actually opened rather than treating a polished deliverable as proof.
Start with Sol for routine drafting, batch work and repeatable agent jobs. Try Astra when the task requires several sources to agree, a template to survive intact or a difficult change to be validated end to end. For codebase research in particular, keep GPT-6.1 Sol in the comparison; the new Sol’s independent regression landed on an everyday task, explaining a repository without dropping relevant facts. Run the same projects with the same tools and score them on the first useful deliverable, then include retries and reviewer time in the bill. That will tell you more about whether Astra is worth the upgrade than another decimal point on a leaderboard.

