· via dev.to (home feed)
OpenAI ships GPT-6.2 Astra and Sol: coding scores nearly tied, prices four times apart
OpenAI made GPT-6.2 Astra and GPT-6.2 Sol generally available on November 17, 2026. Early benchmarks show the pair nearly tied on coding while Astra costs four times more per token.

OpenAI's GPT-6.2 Astra and GPT-6.2 Sol became generally available on November 17, 2026, and the pairing forces a practical decision on anyone running coding agents: pay Astra rates for the hardest reasoning, or default to Sol at a quarter of the token price. A hands-on comparison by developer Taswar Bhatti on dev.to works through the trade-off using Azure pricing, vendor benchmarks and a C# sample you can point at your own tasks.
The naming is asymmetric: GPT-6.2 Astra supersedes GPT-6 Astra, while GPT-6.2 Sol supersedes GPT-6.1 Sol and is built on GPT-6.1.
Identical API surface
According to Microsoft Learn's model reference, both models expose the same shape: IDs gpt-6.2-astra and gpt-6.2-sol dated 2026-11-17, a 1,050,000-token context split as 922,000 input and 128,000 output, text and image in with text out, and support for both the Responses and Chat Completions APIs, with function calling and parallel tool calls reserved for the Responses API. Structured outputs, tools, computer use, reasoning, and reasoning-effort and verbosity controls appear on both, across Global Standard, Data Zone and Global Provisioned Throughput deployments.
One omission to verify yourself: the Learn page no longer lists streaming for either 6.2 model even though it did for both predecessors, so Bhatti recommends testing streaming before depending on it. Because everything else matches, switching between the two in code amounts to changing a deployment name, which is what makes his side-by-side experiment, a .NET 10 program built on Microsoft.Extensions.AI, cheap to reproduce.
What the upgrade claims
Per Microsoft's announcement as relayed in the post, Astra improves on reasoning, writing, coding, computer use and document, slide and spreadsheet generation, with a headline claim of beating GPT-6 Astra's average accuracy while emitting markedly fewer output tokens. Sol's gains target production use: terser technical writing, sharper coding and debugging, better judgement about when to reach for tools, front-end design and a more natural interaction style. No benchmark numbers accompany the announcement, only references to Foundry testing, which Bhatti reads as positioning rather than proof.
A four-times price gap
Under Microsoft's Global Standard short-context pricing, Astra runs $8.00 per million input tokens and $40.00 per million output; Sol runs $2.00 and $10.00, with cached input at $0.40 versus $0.10. That is exactly 4x on input, cached input and output, meaning Astra only matches Sol's bill on a given task if it consumes a quarter of the tokens. Long-context rates are steeper still: $16 in and $60 out for Astra, $4 and $15 for Sol.
Astra did get cheaper against itself: GPT-6 Astra listed $10.00 in, $50.00 out and $1.00 cached, so the new model is 20% cheaper per token and 60% cheaper on cached input. Bhatti's worked example, 1,000 runs consuming 3,000 input tokens (2,000 cached) and 600 output tokens each, comes to $32.80 on Astra against $8.20 on Sol. He also notes that reasoning tokens bill as output, so a terse answer is not automatically a cheap one. Cached input costs 5% of the input price on both models, while cache writes cost 125% of input for the default 30-minute window and 250% for the opt-in 12-hour tier, enabled by setting prompt_cache_options.ttl to 12h in a Responses API request.
Coding close, long-horizon agents apart
OpenAI published launch benchmarks, which DataCamp relays with a vendor-reported label. On FrontierCode 2.0 the two are nearly level at 44.2% for Astra and 43.8% for Sol, with their predecessors at 42.8% and 39.0%, and DataCamp's relayed cost per task on that benchmark is $5.97 for Astra against $1.29 for Sol. The separation shows up on multi-step and security-heavy evaluations: SecureSWE 65.7% to 55.6%, AutomationBench 52.4% to 41.2%, Terminal-Bench-Science 77.1% to 68.0%, Agents' Last Exam 78.3% to 73.0% and GDPval 50.4% to 46.9%. That tracks the stated split: Astra for the hardest work, Sol for everyday production agents.
Caveats worth heeding
The numbers are one vendor's own tests, and several are internal benchmarks nobody outside can reproduce. DataCamp's independent checks also temper the picture: tighter safeguards reportedly produce more refusals, occasionally on legitimate security research, and in one dashboard test the previous Astra handled a messy dataset more sensibly than the new release.
Why it matters
For teams deploying agentic coding systems, this launch is a routing decision more than a model decision. The coding benchmark is effectively a tie while Sol costs roughly a fifth as much per task, which argues for putting routine coding agents on Sol and reserving Astra for long-horizon, security-sensitive work where its lead is far wider. Because the two share an API surface, the choice can be settled empirically per workload, and both Microsoft and Bhatti point out that evaluations on your own tasks, not vendor charts, should make the call. Budget planners should watch the fine print too: reasoning tokens bill as output, and the streaming capability missing from the docs could bite anyone piping responses to users live.
- #openai
- #llms
- #ai-agents
- #azure
- #api-pricing