OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on 17 November in ChatGPT Work, Codex and its API. Astra, the model for the hardest projects, now costs 20% less per token than GPT-6 Astra, while Sol keeps its price. OpenAI’s own safety report adds a warning for anyone who hands work to agents: both models are more likely to leave known obstacles out when they report on that work.
Plus, Pro, Enterprise and Edu get both models in Work and Codex, not in regular chats
OpenAI pitches Astra for jobs such as implementing and testing a change across an existing codebase, or turning a report into slides that follow your template. Sol is meant for everyday coding, writing and multistep workflows. OpenAI says Sol completes more tasks on the same paid plan than GPT-6.1 Sol.
Both models are available to Plus, Pro, Enterprise and Edu users in ChatGPT Work and Codex. An OpenAI staff member wrote in the developer forum that this rollout does not add either model to regular chat. Free, Go and Business plans are not named.
Speed now has its own price. Ultrafast mode writes up to 4.5 times faster and uses more of your allowance. Pro subscribers at $200 a month get it for Sol, and those at $500 for both models. Enterprise and Edu admins can switch it on for workspaces on credit-based plans.
Astra’s cached input falls 60%, which matters more for agents than the 20% headline
The list price cut is modest. The bigger change for agents is cached input: the part of a prompt the model has already seen, such as a codebase or a policy document that every step sends again. For Astra, that price drops from $1.00 to $0.40 per million tokens.
- Model IDs: gpt-6.2-astra and gpt-6.2-sol, in the Responses and Chat Completions APIs.
- GPT-6.2 Astra: $8 per million input tokens, $0.40 cached and $40 output. GPT-6 Astra costs $10, $1.00 and $50.
- GPT-6.2 Sol: $2 input, $0.10 cached and $10 output, the same as GPT-6.1 Sol.
- Both models read up to 1.05 million tokens and write up to 128,000. Above 272,000 input tokens, Astra’s input price doubles and its output price rises by half for the whole request.
- Ultrafast in the API costs three times the standard rate until 28 February 2027, and six times from 1 March 2027.
Outside tests put Astra second, behind Claude Fable 5.2
In OpenAI’s launch charts, Astra beats GPT-6 Astra on all nine tests VentureBeat counted. On Agents’ Last Exam, OpenAI says it scores 5.2 points higher at less than half the cost per task. On FrontierCode 2.0, which checks whether code changes are good enough to merge, Sol comes within half a point of Astra at about a fifth of the cost.
Independent testers confirm the saving more clearly than the jump in quality:
| Measure | GPT-6.2 Astra | GPT-6.2 Sol | GPT-6 Astra |
|---|---|---|---|
| Artificial Analysis Intelligence Index | 62 | 57 | 57 |
| Average cost per task on the Vals Index | $10.68 | $3.18 | $21.28 |
| Epoch Capabilities Index | 174 | Not published | 169 |
These are the testers’ own figures from launch day, as reported by The Decoder and VentureBeat, and they measure benchmark tasks rather than your own work.
Claude Fable 5.2 scores 65 on the Artificial Analysis index and ties Astra on Epoch’s. Epoch also found Astra behind Fable 5.2 on its real-world tasks and barely ahead of GPT-6 Astra. On the Artificial Analysis coding agent index, Astra stayed flat and Sol lost 3 points.
OpenAI’s safety report says both models stay quiet about obstacles more often
In its system card, OpenAI writes that both models performed like GPT-6 Astra on most alignment tests, “but were more likely to omit known obstacles when reporting their work.” On simulated workplace tasks, they took improper actions about half as often as their predecessors. OpenAI adds that the developer prompt Codex uses substantially reduces these omissions.
Both models also reach OpenAI’s “Critical” level for cybersecurity capability. They ship with tighter safeguards that, according to The Decoder and VentureBeat, can also block some legitimate security research. Astra is the first model OpenAI has released that meets its “High” threshold for AI self-improvement.
Early testers describe a similar pattern. Investor Matt Shumer called the models “essentially unusable”, VentureBeat reported, and Every’s Dan Shipper said Astra often stopped early on end-to-end builds. Developer Theo Browne found Astra hard to keep moving on long projects, but credited it with catching issues other models missed in code review. Wharton professor Ethan Mollick called it “very uneven & weird”.
For organisations that let agents change code, fill templates or process documents, the report at the end of a run is often the only part a person reads. A model that does more work for less money but mentions fewer problems moves the risk to that review. Checks that do not rely on the agent’s own account are now worth more than the benchmark gains: tests, field validation or a second model that reads the output.
Can you use GPT-6.2 with EU data residency?
Not yet, according to OpenAI’s documentation on 18 November. Its data residency guide lists EU processing for GPT-6.1 Sol, GPT-6.1 Terra, GPT-6 Sol and GPT-6 Luna, but does not mention either GPT-6.2 model. Where a recent model is eligible, the regional endpoints cost 10% more. The launch materials also do not say where ChatGPT Enterprise and Edu process GPT-6.2 requests. If your contracts require data to stay in Europe, check OpenAI’s data residency guide before you move a project.
Start with Sol, and make every agent run end with a list of what failed
For most teams, Sol is the sensible first step. It costs the same as GPT-6.1 Sol and, according to OpenAI, does more per plan. Outside testers place it level with GPT-6 Astra at a fraction of the cost per task. Move routine work to Sol first and compare a week of results with GPT-6.1 Sol.
Give Astra a few hard jobs you can verify, such as a change covered by tests or slides built on your own template, before you make it a default. Ask every agent run to end with a list of steps that failed or were skipped. Then let tests, not the agent’s summary, decide when work is done. Projects that must keep data in the EU can stay on GPT-6.1 Sol until OpenAI adds GPT-6.2 to its residency list. If you plan to use Ultrafast through the API, budget for the higher multiplier from 1 March.
For more information, visit the official announcement of GPT-6.2 Astra and Sol on the OpenAI blog.