gpt-6.2-astra.
The context window holds up to 922,000 input tokens and 128,000 output tokens. The provider accepts text and image inputs and returns text. Audio and video are not supported. Valid reasoning effort levels are
low, medium, high, xhigh, and max: the API rejects none.
When to use Astra and when to use Sol
OpenAI positions GPT-6.2 Sol as the default for most technical and non-technical tasks, and reserves Astra for deliverables that need the highest quality ceiling.What OpenAI measured
OpenAI highlights cleaner professional deliverables, lower output token usage than GPT-6 Astra, and gains across legal, financial, scientific, and visual benchmarks.- On Agents’ Last Exam, which tests multistep professional workflows, GPT-6.2 Astra scored 5.2 points higher than GPT-6 Astra at under one-third the cost per task, and 5.5 points ahead of Claude Fable 5.2.
- On GDPval v2, OpenAI’s internal evaluation of long-horizon projects in finance, consulting, and data science, it gained 6.2 percentage points over GPT-6 Astra while using 22% fewer output tokens.
- On AA-AnalystAgent, covering quantitative analysis over real spreadsheets and reports, it answered 62.5% of questions right across all five independent runs.
- On Harvey LAB-AA v1.1, which awards credit only when a legal deliverable meets every requirement with no material hallucinations, it more than doubled Fable 5.2’s pass rate.
- On Terminal-Bench-Science 0.1, it gained 9 percentage points over GPT-6 Astra at less than one-third the cost, and set a new high on the Vals Science Index.
- On AA-3D, where human judges evaluate blinded interactive browser 3D experiences, it ranked first at 1,318 Elo, ahead of Fable 5.2 (1,271) and GPT-6.2 Sol (1,259).
What Artificial Analysis measured
Independent testing breaks down all five effort levels ofgpt-6.2-astra. At max, the model scores 62 on the Intelligence Index, tied with xhigh and second overall behind Claude Fable 5.2 at release.
Across individual Artificial Analysis evaluations,
xhigh reached 78.6% on Terminal-Bench-Science 0.1 and 49.0% on FrontierCode-AA 2.0 (above max at 45.4%). On HLE-Diamond, max reached 72.2%, and Terminal-Bench 4.0 landed at 64.1% on both xhigh and max.
xhigh matched max at 62 on the Intelligence Index in roughly half the time per task (7.6 vs. 13.8 minutes) and nearly one-third the wait for the first answer token. Keep max for hard scientific or formal reasoning where xhigh still falls short.Where it does not lead
- On the overall Artificial Analysis Intelligence Index, 62 trails Claude Fable 5.2 (
max, 65), and Astra also places behind Fable 5.2 on CritPt v2, HLE-Diamond, and AA-Omniscience v1.1 Accuracy. - On the Artificial Analysis Coding Agent Index (Codex harness at
max), GPT-6.2 Astra is flat with GPT-6 Astra. - At
max, time to first answer token averages 207 seconds, and output speed stays around 50 tokens per second. - On AA-Omniscience Hallucination Rate (guessing when the model lacks an isolated fact),
maxrecorded 59%: ground tasks in attached files and instruct the model to state when data is missing. - In two internal OpenAI alignment tests, the model was more likely to omit mentioning blockers such as a missing file or unavailable search tool unless prompted to disclose them.
- The model meets OpenAI’s Critical threshold for cybersecurity and High thresholds for biological capabilities and AI self-improvement, so higher-risk requests face stricter safeguards. Its knowledge cutoff is May 18, 2026.
Where it creates more value
Workbooks, memos, and template-driven decks
The inputs are financial statements, contracts, or an existing slide deck. Astra reconciles figures across sections, matches layout and typography, and avoids generic filler. Expected result: a reconciled workbook, a source-backed brief, or a deck that needs fewer review rounds.Complex code changes and interactive frontends
The change spans an existing repository and test suite, or turns a specification into an interactive web app, simulation, or 3D scene. Astra plans long trajectories more cleanly and writes tighter patches. Expected result: a focused diff, passing tests, and a working interface without stray edits in unrelated files.Scientific research and quantitative workflows
The task combines experimental data, mathematical modeling, and terminal scripts, where Astra opened a wide lead over the prior generation. Expected result: tested hypotheses, reproducible analysis scripts, and a clear split between observed data and inference.How to apply it in Tess
1
Move to Astra when Sol falls short
Keep GPT-6.2 Sol for everyday runs. Choose GPT-6.2 Astra for executive deliverables, dense legal or financial work, rich frontends, and scientific workflows.
2
Supply the template and the definition of done
Attach the reference spreadsheet, slide deck, or memo outline. State which numbers must reconcile and what check closes the task.
3
Require explicit blocker reporting
Tell the agent to report any missing file, permission issue, or empty tool result right away instead of working around it silently.
4
Step up to xhigh before max
Use
medium for fast drafts, high for dense code and documents, and xhigh for demanding work. Reserve max for cases that xhigh still misses.5
Keep the prompt prefix stable
Place fixed instructions and reference material at the start of the thread and append new turns at the end so repeated context can be reused cleanly.
Three playbooks to start
1. Reconciled financial model or memo
Input: financial statements, operating assumptions, and target layout.Tools: file reading and spreadsheet or document generation.
2. Slide deck that follows a template
Input: template file or screenshot, project data, and target audience.Tools: image/file reading and slide generation.
3. Code change with blocker disclosure
Input: repository context, expected behavior, architectural constraints, and test command.Tools: code access, terminal, and test runner.
Limits and good practice
- The API rejects
reasoning.effort="none";lowis the minimum. When routed through Amazon Bedrock, sampling parameters such astemperatureandtop_pare rejected. - A 1,050,000-token window (up to 922,000 input tokens) does not replace selection. On the provider API, requests above 272,000 input tokens are billed at higher long-context rates across the entire request.
- OpenAI offers Ultrafast and Fast processing modes on its own API; this page does not claim those modes appear in the Tess picker. Tess credit rates appear in Models and Costs once the slug is in the catalog.
- Keep human approval in place before sending external messages, mutating production data, or acting on financial, legal, or clinical output.

