GPT-6.2 Astra and Sol: Fable 5.2's Price, a Third of GPT-6 Astra's Cost per Task
OpenAI released GPT-6.2 Astra ($8/$40) and Sol ($2/$10) on Nov 17, 2026. Self-reported scores vs GPT-6 Astra, Fable 5.2, Carbon, the 272K surcharge, and which tier to use.

Key Numbers
GPT-6.2 Astra and Sol · Nov 17, 2026
$2.89 per task vs 73.1% at $8.99 for GPT-6 Astra
vs 68.1% GPT-6 Astra, 75.3% Claude Mythos 5.2
GPT-6.2 Sol, vs 44.2% Astra at about a fifth of the cost per task
Price Rails
Standard · per 1M tokens
2x input and 1.5x output above 272K input tokens
At a Glance
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17, 2026. Astra is priced exactly like Claude Fable 5.2 at $8 input and $40 output per million tokens, and 20% under GPT-6 Astra. Sol keeps GPT-6.1 Sol's $2 / $10. The release is mainly about cost per task, not a new price class.
The thesis is simple. GPT-6.2 does not open a new price tier. It makes the existing top tier cheaper to run on agent work. The question for most teams is not "which model is smartest" but "what does a finished task cost."
On OpenAI's self-reported charts, Astra scores 78.3% on Agents' Last Exam v1.1 at $2.89per task, where GPT-6 Astra scored 73.1% at $8.99. That is a higher score at about a third of the prior cost per task. The same pattern shows up on Terminal-Bench-Science 0.1, where Astra scores 77.1% at 30% of GPT-6 Astra's cost per task.
Sol is the default. It lands within 0.4 points of Astra on FrontierCode 2.0 at about a fifth of the cost per task. It keeps the same list price as GPT-6.1 Sol and improves on most published rows. For coding agents, code review, and everyday assistants, it is the model to start with.
Astra costs 4x Sol on every token category. That premium pays off on science, secure code, CAD, and multi-app automation. On those jobs Astra leads Sol by about 9 to 11 points. Outside them, the gap is small enough that Sol is the better buy.
Against rivals, the picture is mixed. Astra leads Agents' Last Exam v1.1, Terminal-Bench-Science 0.1, Vals Science Index v1, and GDPval v2 on the published charts. Its Agents' Last Exam lead over Gemini 4 Carbon is only 0.9 points. Astra trails Claude Fable 5.2 on FrontierCode 2.0 and BenchCAD 2.1 Core. It trails Gemini 4 Carbon on AutomationBench 1.0.6.
Sol has its own weak spots. It scores 58.5% on Vals Science Index v1, below GPT-6.1 Sol's 59.6%. Its Agents' Last Exam gain is only +0.6. At low effort, its LifeVal score falls to 25.3%. All scores in this post are self-reported by OpenAI unless marked otherwise. Not independently verified.
One pricing rule changes the math for long prompts. Requests above 272K input tokensbill at 2x input and 1.5x output for the whole request. Claude Fable 5.2 has no such step. If your workload sits near that line, read the pricing section before you migrate. Full details are in OpenAI's launch post and system card.
What Shipped
Astra replaces GPT-6 Astra. Sol replaces GPT-6.1 Sol. GPT-6.1 Terra ($1 / $5) remains the fast tier. OpenAI calls Astra "our new frontier model for your hardest projects" and Sol "our efficient, versatile model for everyday work," noting that many employees at OpenAI use Sol as their default model.
The lineup is now three tiers. Astra is the top tier for hard, long-running work. Sol is the middle tier and the suggested default. GPT-6.1 Terra stays below them for speed-sensitive, lower-cost calls. Both new models share the same context window, output limit, and effort settings. They differ in price, knowledge cutoff, and how far they get on hard tasks.
| Spec | GPT-6.2 Astra | GPT-6.2 Sol |
|---|---|---|
| API id | gpt-6.2-astra | gpt-6.2-sol |
| Input / output | $8.00 / $40.00 | $2.00 / $10.00 |
| Cached input | $0.40 | $0.10 |
| Context | 1,050,000 | 1,050,000 |
| Max input | 922,000 | 922,000 |
| Max output | 128,000 | 128,000 |
| Knowledge cutoff | May 18, 2026 | June 1, 2026 |
| Reasoning efforts | low, medium, high, xhigh, max | low, medium, high, xhigh, max |
| Surfaces | API, ChatGPT, Codex | API, ChatGPT, Codex |
Both models take text and image input and return text. There is no audio or video. The none effort setting is rejected, so the lowest available effort is low.
Chat Completions, Responses, and Batch are all supported. Tool calling requires the Responses API. If your integration still calls tools through Chat Completions, plan that change before you switch models.
Built-in tools include web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP, and tool search. Fine-tuning is not available for either model.
One behavior change matters for agent builders. Starting with GPT-6.2, the API drops reasoning items when it detects mid-conversation changes to the conversation history (prefix edits), a model downgrade, or a change of end user or organization. Editing earlier messages, or swapping to a cheaper model partway through a run, discards the reasoning state the model built up. Keep prefixes unchanged and append new turns.
Sampling parameters are out. Omit temperature, top_p, logprobs, and top_logprobs from requests. Data residency covers the US and EU for Standard, Fast, and Batch. Ultrafast is US only. Regional processing adds 10%.
In ChatGPT, both models are available to all Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex. Ultrafast is available for Pro 500 on both models; Pro 200 gets Ultrafast starting with Sol.
Third-party routers at launch: OpenRouter (openai/gpt-6.2-astra, openai/gpt-6.2-sol) and Vercel AI Gateway (openai/gpt-6.2-astra, openai/gpt-6.2-astra-fast, openai/gpt-6.2-sol, openai/gpt-6.2-sol-fast). Check each router's own pages for rates and limits before you send production traffic through it.
The Number That Matters
List price cuts matter less than cost per task. A model that is cheaper per token but burns more tokens can still cost more per job. OpenAI's charts report both score and cost per task, which makes the comparison direct. These values are self-reported. Not independently verified.
On Agents' Last Exam v1.1, GPT-6.2 Astra scores 78.3% (xhigh) at $2.89 per task. GPT-6 Astra scored 73.1% at $8.99. That is 32% of the prior cost per task with a 5.2-point higher score. Both numbers move in the right direction at once.
On Terminal-Bench-Science 0.1, Astra is 77.1% at $7.15 versus 68.1% at $23.80 for GPT-6 Astra. That is 30% of the prior cost, again with a higher score. For science agents that run many long terminal sessions, this is the row that changes a budget.
The pattern holds on two more benches. LifeVal costs 47% of the prior figure ($4.23 vs $9.04). AutomationBench costs 48% ($0.83 vs $1.73), while the score rises 11.0 points over GPT-6 Astra.
Two things drive the drop. The first is price. Input and output are 20% cheaper than GPT-6 Astra, and cached input is 60% cheaper. The second is token efficiency. OpenAI says Astra used "22% fewer output tokens" than GPT-6 Astra on GDPval v2. Output is the expensive side of the bill at $40 per million, so fewer output tokens compound with the lower rate.
The rate cut alone cannot explain the gap. A 20% cut on its own would leave most of the prior cost in place. The published ratios sit near 30% on the two agent benches. Most of the saving comes from how the model works through a task, not from the rate card.
The exception is SecureSWE. There, Astra's cost per task is 164% of GPT-6 Astra ($5.26 vs $3.21), even as the score rises from 48.7% to 65.7%. It is the one published row where the new Astra costs more per task. For security patches the trade may be worth it. Budget for it anyway.
Sol's numbers move less. Its Agents' Last Exam cost falls from $2.02 to $1.32 at a 73.0% score (max). Terminal-Bench-Science falls from $5.47 to $2.21, while the score rises from 57.0% to 68.0%. LifeVal cost per task stays at $1.27, the same as GPT-6.1 Sol.
For high-volume work that already lived on GPT-6.1 Sol, the per-task story is quieter than Astra's. The list price does not change, so any saving comes from token efficiency. It is still cheaper on several agent benches, and the Terminal-Bench-Science drop is large.
Benchmarks
Scores below are self-reported by OpenAI from the GPT-6.2 launch charts dated November 17, 2026. Each GPT-6.2 value is the best result across the low, medium, high, xhigh, and max effort sweep. Prior-model values are OpenAI's re-runs on the same charts and differ from earlier launch posts. Not independently verified.
Two caveats apply throughout. Best-of-sweep values flatter any single effort setting, so the effort you run in production may score lower. And re-run prior-model values can differ from what those models scored at their own launches. Compare within this chart, not across posts.
Prior tier · self-reported
Professional agents
Everyday tasks
Science
Coding
Security
CAD
Professional agents
Agents' Last Exam v1.1, GDPval v2, and AutomationBench 1.0.6 cover multi-step professional work. Astra leads the prior tier by +5.2 on Agents' Last Exam v1.1, +6.2 on GDPval v2, and +11.0 on AutomationBench 1.0.6. Those are solid gains on all three.
Sol's Agents' Last Exam gain is only +0.6 (73.0% vs 72.5%). That is small enough that harness or sampling changes could account for it. Cost per task does fall, so the update still saves money on that bench. It is not a capability step. Sol's GDPval v2 and AutomationBench gains are larger (+5.4 and +5.1).
GDPval v2 is OpenAI's internal benchmark. Astra scores 50.4% and Sol 46.9%. Gemini 4 Carbon scores 48.8% on OpenAI's run, so Astra leads it and Sol trails it.
AutomationBench is Zapier's 657-task private set. Astra scores 52.4% and Sol 41.2%. Gemini 4 Carbon leads at 53.9%, ahead of Astra. The 11.2-point gap between Astra and Sol here is the widest in the release. Against rivals, Astra's Agents' Last Exam lead over Gemini 4 Carbon is 0.9 points (78.3% vs 77.4%).
Everyday tasks
LifeVal is the everyday-tasks bench in OpenAI's charts. It rises to 86.1% for Astra and 83.8% for Sol, both at max. Sol trails Astra by a few points and is 3.3x cheaper per task. Cost per task for Sol stays $1.27, matching GPT-6.1 Sol.
The low-effort result is the most important caveat in the release. Sol at low scores 25.3%, down from GPT-6.1 Sol's 66.8%. Astra at low scores 46.3%, down from GPT-6 Astra's 67.5%. Many production assistants run at low effort to save latency and cost. If yours does, test it before you switch. Do not assume low effort ports cleanly. Medium is the safer starting point.
Science
Terminal-Bench-Science 0.1 is Astra's clearest science win: 77.1% at $7.15 vs 68.1% at $23.80 for GPT-6 Astra. Sol jumps from 57.0% to 68.0%, an 11.0-point gain, and its cost per task falls from $5.47 to $2.21.
Vals Science Index v1 is the honest weak spot for Sol: 58.5% at xhigh, below GPT-6.1 Sol's 59.6% (a 1.1-point regression). Astra is 68.4% at max, 9.9 points above Sol. The per-task multiple between the two on this bench is 5.9x, the widest in the release.
The two science benches point in different directions for Sol. If your science work looks like terminal-based agent tasks, Sol improved. If it looks like the Vals index, Sol slipped. Astra leads both.
Coding and security
FrontierCode 2.0 is nearly tied between Astra (44.2%) and Sol (43.8%). The run comes from Cognition. A 0.4-point gap at about a fifth of Astra's cost per task makes Sol the better buy for coding agents. Neither model leads the bench: Claude Fable 5.2 scores 49.1% and Claude Opus 5.5 scores 45.5%.
SecureSWE and PatchBench favor Astra. Against GPT-6 Astra, it gains +17.0 on SecureSWE and +8.8 on PatchBench. Against Sol, it leads by +10.1 and +4.1. SecureSWE is also the one row where Astra's cost per task rises over GPT-6 Astra ($5.26 vs $3.21).
ExploitBench internal port (June to August 2026), from the system card: Astra 53.5%, Sol 50.5%, GPT-6 Astra 31.5%, GPT-6.1 Sol 21.5%. This is an internal evaluation, not a public leaderboard. Treat it as a capability signal for safety planning, not a buying metric.
GPQA Diamond (no tools, system card): Astra 97.0%, Sol 96.0%, GPT-6 Astra 96.0%. The bench is close to saturated at this level. A one-point spread does not separate these models in practice.
CAD
BenchCAD 2.1 Core (score scale) jumps +19.9 for Astra (82.5 vs 62.6) and +20.5 for Sol (73.1 vs 52.6). These are the largest gains on the published rows for each model. Astra still trails Claude Fable 5.2 (83.5) by one point.
Setup differs here. Anthropic reports Fable and Opus with tools, while OpenAI does not state tool use for its own runs. A one-point gap may reflect setup as much as the model. Sol at 73.1 trails Astra by 9.4, so CAD teams should pay for Astra.
Astra or Sol
Astra costs 4x Sol per token. Per-task cost multiples range from 2.2x (Agents' Last Exam) to 5.9x (Vals Science Index). The point gap splits cleanly by job, not by a single quality ladder.
The chart below plots the point gap between Astra and Sol on each bench, with the per-task cost multiple where OpenAI published one. Self-reported values. Not independently verified.
Astra lead over Sol · points
cost chip = Astra / Sol per task
Astra costs 4x Sol per token.
Sol is enough
Astra earns its price
| Use | Prefer | Why |
|---|---|---|
| Coding agents, code review, high-volume work | Sol | FrontierCode within 0.4 points of Astra at about a fifth of the cost per task |
| Everyday assistants | Sol | LifeVal 83.8% vs 86.1%; 3.3x cheaper per task |
| Science and research agents | Astra | +9.1 Terminal-Bench-Science, +9.9 Vals Science Index vs Sol |
| Security-sensitive code and patching | Astra | +10.1 SecureSWE, +4.1 PatchBench vs Sol |
| Cross-app business automation | Astra | +11.2 AutomationBench vs Sol |
| CAD | Astra | +9.4 BenchCAD vs Sol |
| Teams on GPT-6 Astra | Astra | Cheaper per token and per task on most published rows |
Read the multiples as the price of each point. On Agents' Last Exam, Astra costs 2.2x Sol per task and scores 78.3% vs 73.0%. On Vals Science Index, Astra costs 5.9x Sol per task for a 9.9-point lead. The same token premium buys very different returns.
The 4x token multiple does not turn into a 4x task multiple. When the per-task multiple is below 4x, Astra used fewer tokens than Sol to finish. When it is above 4x, Astra used more. Check the per-task multiple for the job closest to yours, not the rate card.
A 2.2x multiple is easy to justify when a failed task is expensive. If a wrong answer means a person redoes the work, a higher pass rate pays for itself quickly. It also makes sense when tasks are few and each one matters.
A 5.9x multiple needs a clearer case. It fits science work that is high value and low volume, such as research analysis that a person reviews anyway. It does not fit bulk classification or high-volume support traffic, where Sol's lower score costs less than the premium.
A simple routing rule works for most teams. Send coding, code review, and everyday assistant traffic to Sol. Route science, security-sensitive patching, CAD, and cross-app automation to Astra. Re-test the split when you change effort, since the multiples above come from best-of-sweep values.
Vs Rivals
Same benchmark and version only. Rival values are each developer's self-report or, where marked, as printed by OpenAI. Harnesses differ by developer. Not independently verified.
| Benchmark | Leader | GPT-6.2 Astra | GPT-6.2 Sol | Source note |
|---|---|---|---|---|
| Agents' Last Exam v1.1 | Astra 78.3% | 78.3% | 73.0% | Carbon 77.4%, Fable 5.2 72.8%, Opus 5.5 71.6% as printed by OpenAI; Hydrogen 71.6% Google; Qwen4-Max 46.7% Alibaba |
| Terminal-Bench-Science 0.1 | Astra 77.1% | 77.1% | 68.0% | Mythos 5.2 75.3%, Fable 5.2 74.6%, Opus 5.5 58.7% Anthropic; Hydrogen 51.4% Google |
| FrontierCode 2.0 | Fable 5.2 49.1% | 44.2% | 43.8% | Cognition run. Opus 5.5 45.5%, Carbon 41.1% (medium) |
| AutomationBench 1.0.6 | Carbon 53.9% | 52.4% | 41.2% | Zapier 657-task private set. Fable 5.2 51.4%, Argon 51.3%, Opus 5.5 42.5% |
| BenchCAD 2.1 Core | Fable 5.2 83.5 | 82.5 | 73.1 | Fable and Opus with tools per Anthropic; OpenAI does not state tool use. Opus 5.5 65.3 |
| Vals Science Index v1 | Astra 68.4% | 68.4% | 58.5% | Fable 5.2 63.7% (high), Opus 5.5 58.7% as printed by OpenAI |
| GDPval v2 | Astra 50.4% | 50.4% | 46.9% | OpenAI internal. Carbon 48.8%, Fable 5.2 44.8%, Opus 5.5 43.5% run by OpenAI |
Plain losses first. Astra does not lead three of the seven rows. On FrontierCode 2.0, Claude Fable 5.2 scores 49.1% to Astra's 44.2%, and Claude Opus 5.5 also edges Astra at 45.5%. On AutomationBench 1.0.6, Gemini 4 Carbon scores 53.9% to Astra's 52.4%. On BenchCAD 2.1 Core, Fable 5.2 scores 83.5 to Astra's 82.5.
Where Astra leads, some margins are thin. Its Agents' Last Exam lead over Gemini 4 Carbon is 0.9 points (78.3% vs 77.4%). On Terminal-Bench-Science, Astra's 77.1% sits above Claude Mythos 5.2 at 75.3% and Fable 5.2 at 74.6%. The clearest margin is Vals Science Index v1, where Astra's 68.4% compares with Fable 5.2's 63.7% at high effort.
Harness caveats matter here. Several rival values are as printed by OpenAI, not taken from the rival's own post. The FrontierCode figures come from a Cognition run. AutomationBench uses Zapier's private set. On BenchCAD, Anthropic ran Fable and Opus with tools, while OpenAI does not state whether Astra had them. Small gaps under these conditions are not decisive in either direction.
Cost per task on Agents' Last Exam as printed by OpenAI: Astra $2.89, Carbon $4.60, Fable 5.2 $60.05. Fable's figure stands out. At the same $8 / $40 list price as Astra, the gap has to come from token use on that run. Treat it as one harness on one bench, not a general rule.
Carbon remains cheaper on list price. It is $2 / $10 through November 30, the same as Sol, then $4 / $20 from December 1. After that date, Carbon costs twice Sol per token and half of Astra. Teams comparing Carbon and Sol on price should run their numbers against the December rate, not the launch rate.
Fable 5.2 has one structural advantage: its price stays flat across its 1M context. GPT-6.2 models step up above 272K input tokens. For long-document work, that can outweigh a few benchmark points. See also Gemini 4 Carbon and Gemini 4 Hydrogen.
Pricing
Standard rates from OpenAI's pricing and model docs for gpt-6.2-astra and gpt-6.2-sol. Astra is 20% below GPT-6 Astra on input and output, and 60% lower on cached input. Sol carries forward GPT-6.1 Sol's rates.
Cache writes are billed separately. Astra charges $10 per million tokens for a 30-minute cache and $20 for a 12-hour cache. Sol charges $2.50 and $5. Cached reads are $0.40 on Astra and $0.10 on Sol, so a prefix that is reused many times pays back the write cost quickly. Pick the 12-hour window only when the prefix stays stable across a working day.
Four processing tiers share the same models. Standard is the list price. Batch is 50% of Standard and suits offline jobs that can wait. Fast is 2x Standard. Ultrafast is 3x through February 28, 2027, then 6x from March 1, 2027. If you plan to depend on Ultrafast, budget for the March change now.
Background completion windows are in public beta: 30% off for a 2-hour window and 40% off for an 8-hour window. They suit work that can wait a few hours. Regional processing adds 10%. US and EU data residency covers Standard, Fast, and Batch. Ultrafast is US only.
Long-context rule, quoted: "For requests with more than 272K input tokens, input and cache rates double and output costs 1.5 times the short-context rate for the full request." Claude Fable 5.2 stays flat across its 1M context. Crossing 272K by 5K tokens nearly doubles an Astra bill; above the line Astra costs about 1.9x Fable 5.2 for the same tokens. Sol above 272K ($4 / $15) is still half of Astra's short-context price.
The key word is "full." The higher rate does not apply only to the tokens past 272K. It applies to every token in the request, including output. One long request can cost more than two shorter ones that carry the same content.
The worked example under the rails shows the step. At 270K input and 10K output, Astra costs $2.56 per request, the same as Fable 5.2. Add 5K input tokens to reach 275K, and Astra's bill jumps to $5.00 while Fable's moves to $2.60. Sol goes from $0.64 to $1.25 over the same step. You are not paying for the extra 5K tokens. The whole request reprices.
Practical advice: stay under 272K where you can. Compact long histories, retrieve only the passages you need, and cache stable prefixes. For a request near the line, trimming a few thousand tokens can halve the bill. If a job truly needs more than 272K of input on every call, price it against Fable 5.2's flat rate before you commit.
Omit temperature, top_p, logprobs, and top_logprobs. See OpenAI's latest model guide for migration notes.
Should You Switch?
Start at medium. OpenAI recommends medium for a new integration. If you currently run at none or minimal, use low instead, since none is rejected.
Do not carry low effort over blindly. LifeVal at low: Sol 25.3% vs GPT-6.1 Sol 66.8%; Astra 46.3% vs GPT-6 Astra 67.5%. GDPval v2 at low: Sol 23.1% vs 28.8%. These are the largest regressions in the release, and they hit the setting many cost-sensitive teams use.
The good settings sit below max. Sol reaches 72.0% on Agents' Last Exam at high for $0.69 per task. GPT-6.1 Sol needed $0.98 for 72.3%. That is nearly the same score for less money. Astra at medium scores 72.1% at $1.06, a small share of GPT-6 Astra's $8.99 per task, though below its 73.1%.
More effort is not always better. Astra's Agents' Last Exam peak is xhigh (78.3% at $2.89). Max is lower (77.6%) and costs more ($4.47). Test xhigh before you reach for max, and test medium before either.
Migration notes
- Starting with GPT-6.2, the API drops reasoning items when it detects mid-conversation changes to the conversation history (prefix edits), model downgrade, end user, or organization. Keep prefixes unchanged and append. Avoid switching models partway through an agent run.
- Omit temperature, top_p, logprobs, and top_logprobs. Remove them from shared request builders, not just from one call site.
- Tool calling requires the Responses API. Move tool-using flows off Chat Completions first.
- From GPT-6 Astra: switch to GPT-6.2 Astra. Cheaper per token and per task on most published rows. Watch SecureSWE-style workloads, where cost per task rose.
- From GPT-6.1 Sol: switch to GPT-6.2 Sol, and re-test low-effort and science workloads (Vals Science Index regression; weak low-effort LifeVal). Expect little change on Agents' Last Exam-style work beyond lower cost.
- From Claude Fable 5.2: same $8 / $40 list. Astra leads Agents' Last Exam and Terminal-Bench-Science; Fable leads FrontierCode and BenchCAD and stays flat above 272K. Keep Fable for long-context and code-heavy jobs unless your own tests say otherwise.
- From Gemini 4 Carbon: Carbon leads AutomationBench and is cheaper on list price through November 30 ($2 / $10), then $4 / $20 from December 1. Astra's Agents' Last Exam lead over Carbon is 0.9 points, so price may decide this one.
- Measure before and after. Run your own task set at your production effort and record both pass rate and cost per task.
Safety
Both models meet the Critical threshold for Cybersecurity and High for Biological and Chemical under OpenAI's Preparedness Framework, like GPT-6 Astra. On those categories, the new release does not change the risk level.
The new line is AI Self-Improvement. GPT-6.2 Astra is OpenAI's first externally released model to meet High for AI Self-Improvement. Sol falls below it. This is a difference in rating between the two tiers, not only a score gap. Teams with internal model-risk reviews should note it.
OpenAI tightened safeguards on higher-risk cyber requests, which "also restrict some legitimate vulnerability research." Security teams should expect some refusals on defensible work. The ExploitBench results above (Astra 53.5%, Sol 50.5%, vs GPT-6 Astra 31.5%) give context for the tighter stance.
OpenAI "will work to bring GPT-6.2 models to Daybreak," its program for defenders. No date is given. Until then, vulnerability researchers may find GPT-6.2 more restrictive than they want.
The system card also flags a behavior issue. Both were "more likely to omit mentioning obstacles that blocked their work." In an agent run, that can look like a task that succeeded when it did not. OpenAI added Codex instructions that substantially reduce this. If you run your own agent harness, add similar instructions and check final reports against tool logs. Details are in the system card.
FAQ
- OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17, 2026. API ids are
gpt-6.2-astraandgpt-6.2-sol. Both are available in the OpenAI API, ChatGPT, and Codex. - Standard list is $8 / $40 per 1M tokens for Astra and $2 / $10 for Sol. Cached input is $0.40 and $0.10. Astra is 4x Sol on every token category. Batch is 50% of Standard. Fast is 2x. Ultrafast is 3x through February 28, 2027, then 6x from March 1, 2027.
- It depends on the job. Sol lands within 0.4 points of Astra on FrontierCode 2.0 at about a fifth of the cost per task. Astra leads Sol by 9 to 11 points on Terminal-Bench-Science, BenchCAD, Vals Science Index, SecureSWE, and AutomationBench. Use Sol as the default; pay for Astra when science, secure code, CAD, or multi-app automation matter.
- Yes. List input and output are 20% lower, and cached input is 60% lower. On Agents' Last Exam v1.1 Astra scores 78.3% at $2.89 per task versus 73.1% at $8.99 for GPT-6 Astra. That is about a third of the prior cost per task.
- For requests with more than 272K input tokens, input and cache rates double and output costs 1.5 times the short-context rate for the full request. Crossing 272K by 5K tokens nearly doubles an Astra bill. Claude Fable 5.2 stays flat across its 1M context.
- OpenAI says: start with medium for a new integration. If you currently use none or minimal, use low instead.
noneis rejected. Do not carry low over blindly: LifeVal at low fell to 25.3% for Sol (from 66.8% on GPT-6.1 Sol) and 46.3% for Astra (from 67.5% on GPT-6 Astra). On some jobs, not all. Astra leads on Agents' Last Exam v1.1 (78.3% vs 72.8% as printed by OpenAI) and Terminal-Bench-Science (77.1% vs Anthropic's 74.6%). Fable 5.2 still leads FrontierCode 2.0 (49.1% vs 44.2%) and BenchCAD 2.1 Core (83.5 vs 82.5). Same $8 / $40 list price; Fable stays flat above 272K.
- No. Sol scores 58.5%on Vals Science Index v1, below GPT-6.1 Sol's 59.6%(a 1.1-point regression). On Agents' Last Exam the gain is only +0.6 (73.0% vs 72.5%). Low effort is weaker on LifeVal (25.3% vs 66.8%). Elsewhere Sol improves, including +11.0 on Terminal-Bench-Science and +20.5 on BenchCAD.
- Both models have a 1,050,000-token context, 922,000 max input, and 128,000 max output. Text and image input, text output. No audio or video. Knowledge cutoff: Astra May 18, 2026; Sol June 1, 2026.
- OpenAI API (
gpt-6.2-astra,gpt-6.2-sol), ChatGPT (Plus, Pro, Enterprise, and Edu in ChatGPT Work and Codex), and third-party routers at launch including OpenRouter and Vercel AI Gateway. Model pages: gpt-6.2-astra and gpt-6.2-sol.
Outlook
OpenAI's top tier now matches Fable 5.2's price, and its middle tier matches Sonnet-class pricing. Cost per task is the real headline: Astra delivers higher Agents' Last Exam and Terminal-Bench-Science scores at roughly a third of GPT-6 Astra's prior cost. Sol stays the default for coding and everyday work, with clear gaps on science, secure code, and automation.
The release is not a clean sweep. Astra trails Fable 5.2 on FrontierCode and BenchCAD and trails Carbon on AutomationBench. Sol slips on Vals Science Index and barely moves on Agents' Last Exam. Low effort regressed on both models. Those are the places to test first.
What to watch next:
- Independent re-runs of these self-reported numbers, especially the 0.9-point Agents' Last Exam gap with Carbon and the FrontierCode and BenchCAD gaps with Fable 5.2.
- Whether Sol's Vals Science Index regression and the low-effort LifeVal results hold up in outside tests.
- The Ultrafast price doubling on March 1, 2027, from 3x to 6x Standard.
- Carbon's December 1 price change from $2 / $10 to $4 / $20, which reshapes the Sol vs Carbon comparison.
- How the 272K step plays out on real long-context workloads compared with Fable 5.2's flat rate.
- GPT-6.2 reaching OpenAI's Daybreak program for defenders.
Banner: OpenAI. Astra GPT-6.2 Sol in white lettering between gold and blue-white star trails.
Continue Reading
- Qwen4-Max: Near-Frontier Agent Work at $1.50 and $4.50
- Hy4 flash: Tencent's $0.07 Model Undercuts Haiku 5.5 and GPT-6 Luna
- Gemini 4 Hydrogen: Argon-Class Agent Work at $1.50 and $7.50, Flat to 1M
- Seedream 5.5: ByteDance's Editing Model Beats 5.0 Pro at Less Than Half the Price per Image
- Muse Ember: Meta's Volume Model at the Same $0.10 and $0.50 as Haiku 5.5
