Key points
- Two distinct roles: OpenAI released
gpt-6.2-astraandgpt-6.2-solon November 17, 2026. Astra targets complex professional work, science, computer use and difficult engineering; Sol handles everyday coding, writing and multistep agents - Live on APIYI: both models are available in the Default, SVIP and CodexResponses groups through the Responses API
- Astra costs less; Sol keeps its price: Astra is $8/$40 per 1M input/output tokens, down 20% from GPT-6 Astra, with cached input down 60%. Sol retains GPT-6.1 Sol’s $2/$10 pricing
- Shared capacity: both provide a 1,050,000-token context window, 922,000 maximum input tokens and 128,000 maximum output tokens, with text and image input and text output
- Efficiency is central: on OpenAI’s GDPval v2, Astra gains 6.2 percentage points over GPT-6 Astra while using 22% fewer output tokens; Sol gains 5.3 points over GPT-6.1 Sol
Background
Rather than forcing one model to cover every workload, GPT-6.2 separates the hardest projects from everyday work. Astra handles deep reasoning, cross-tool execution and professional deliverables. Sol brings near-flagship capability to a lower price point. The release emphasizes output quality and token efficiency together. Both models aim to complete difficult tasks with fewer output tokens. For API users, the upgrade should therefore be measured by task completion and total tokens consumed, not list price alone. APIYI connected both models on November 18, 2026. Standard-tier model prices match OpenAI’s list prices; savings come through recharge bonuses rather than a modified token rate.Deep dive
Choosing between Astra and Sol
GPT-6.2 Astra
Built for cross-service debugging, complex code changes, scientific analysis, professional decks and spreadsheets, precise computer use and interactive 3D design. Its priority is completion quality on the hardest tasks
GPT-6.2 Sol
Built for pull request reviews, revisions based on feedback, meeting-note synthesis, everyday writing and agents at scale. Its priority is the balance of intelligence, speed and task cost
Official benchmark results
These figures come from OpenAI’s November 17, 2026 launch material and were checked on November 18. Some evaluations were designed by third parties such as Artificial Analysis and Cognition, but the figures quoted here are OpenAI’s reported results and have not been independently reproduced by APIYI.
Specifications
OpenAI’s model documentation lists Responses tools including web search, file search, code interpreter, hosted shell, computer use, MCP and multi-agent delegation. New projects that need a tool chain should start with
/v1/responses rather than designing around a text-only chat flow.
Practical use
Recommended scenarios
- Cross-service incident diagnosis: give Astra the failing workflow, code and logs, then ask it to implement a fix and run the relevant tests
- Everyday code review: use Sol to review a pull request, revise the implementation from feedback and add targeted tests
- Professional knowledge work: use Astra to turn financial reports, templates and domain conventions into workbooks, memos or slide decks
- Agents at scale: measure Sol’s cost per completed task first and reserve Astra for complex steps where failure is expensive
- Long documents and repositories: use the 1.05M context to reduce splitting, while accounting for the pricing tier above 272K input tokens
Responses API example
Best practices
- Preserve stable prefixes: keep system instructions, tool definitions and earlier input intact, and append new instructions to improve cache reuse and preserve reasoning across turns
- Do not send
none: both models rejectreasoning_effort: none; check this parameter when migrating from an older model - Watch the context tier: a request above 272K input tokens moves the entire request into long-context billing, so do not fill the 1.05M window by default
- Measure total task cost: track completion rate, output tokens, cache hits and manual rework alongside token prices
- Require explicit obstacle reporting: OpenAI’s alignment evaluations found that both models are more likely to omit blockers such as missing files or unavailable tools. Critical agent workflows should require a clear blocker list
Pricing and availability
Standard-tier pricing
These rates apply to Standard requests with no more than 272K input tokens, in US dollars per 1M tokens:Model prices match the provider and may change when provider pricing changes. This table is for reference; the Model Pricing tab in the top navigation is authoritative: Model Pricing.
Groups and recharge promotions
Both models are available in Default, SVIP and CodexResponses. Model rates match OpenAI’s Standard tier, and savings come through recharge bonuses: public recharge tiers bring the effective cost to roughly 83–91% of the list-price amount, depending on the single-recharge tier, without changing the model amount shown in usage logs. 📖 See recharge promotions and effective-cost calculationsSummary and recommendations
In short, Astra is stronger and cheaper than its predecessor, while Sol moves closer to flagship capability without changing its standard price. Start withgpt-6.2-sol for everyday coding, writing and agents at scale. Use gpt-6.2-astra for complex professional deliverables, science and long-horizon engineering, and validate the benefit using total tokens and completion rate on your own tasks.
Sources: OpenAI’s launch announcement
openai.com/index/introducing-gpt-6-2-astra-sol/ (November 17, 2026), the official GPT-6.2 Astra and Sol model pages, and OpenAI’s Codex release notes, checked November 18. APIYI group availability and Standard-tier pricing come from the live platform pricing API on November 18, 2026.