GPT-6.2 Astra and GPT-6.2 Sol are now available on Bifrost
Akshay Deo
Nov 17, 2026 · 4 min read

OpenAI shipped GPT-6.2 Astra and GPT-6.2 Sol today, and Bifrost supports both from day 0. Astra is OpenAI's frontier model for the hardest work, now 20% cheaper than GPT-6 Astra. Sol is the efficient everyday model, same rate card as GPT-6.1 Sol, but it finishes the same work on fewer tokens.
If you already route OpenAI traffic through Bifrost, switching is one line: change the model string to openai/gpt-6.2-astra or openai/gpt-6.2-sol. Your routing rules, fallbacks, virtual keys, budgets, guardrails and logs carry over unchanged.
Why GPT-6.2 Astra and Sol
OpenAI pitches the pair as frontier performance at a cost that works at scale. Astra takes the hard, long-running jobs. Sol takes the volume. All figures below are OpenAI-reported unless noted.
GPT-6.2 Astra
- Cheaper frontier. $8 / $40 per million input / output tokens, down 20% from GPT-6 Astra. Cached input drops from $1 to $0.40.
- Better at long agent work. On Agents' Last Exam, Astra at
xhighbeats GPT-6 Astra atmaxby 5.2 points, at under half the cost per task. - Professional work on fewer tokens. +6.2 points on GDPval v2 with 22% fewer output tokens.
- Science and terminal tasks. +9 points on Terminal-Bench-Science at under a third of the cost.
- Codex results. OpenAI says early testers saw cleaner code style, more focused tests and fewer revision rounds in existing projects.
- Independent check. Artificial Analysis ranks Astra second on its Intelligence Index, behind Claude Fable 5.2 and five points above GPT-6 Astra.
GPT-6.2 Sol
- Same price, less work. $2 / $10, identical to GPT-6.1 Sol. The saving comes from token efficiency.
- Close to Astra on code. On FrontierCode 2.0, Sol nearly matches Astra on merge-ready changes at about a fifth of the cost.
- GPT-6 Astra-level intelligence for less. Artificial Analysis scores Sol at 57, matching GPT-6 Astra at about 15% of its cost per task.
Suggested setup: Astra on high or xhigh for planning, hard coding and research. Sol for subagents, chat, extraction and high-volume calls. Bifrost lets you run both behind one virtual key and route by workload.
Pricing
Astra gets a 20% cut on input and output and a 60% cut on cached input. Sol keeps GPT-6.1 Sol's rates.
Model | Bifrost model string | Input ($/1M) | Cached input ($/1M) | Output ($/1M) |
|---|---|---|---|---|
GPT-6.2 Astra |
| $8.00 | $0.40 | $40.00 |
GPT-6 Astra (previous) |
| $10.00 | $1.00 | $50.00 |
GPT-6.2 Sol |
| $2.00 | $0.10 | $10.00 |
GPT-6.1 Sol (previous) |
| $2.00 | $0.10 | $10.00 |
Watch the long-context tier. Sol bills the whole request at $4 / $0.20 / $15 once input crosses 272K tokens, and the GPT-6 Astra rate card had the same rule. Bifrost logs the per-request cost, so you can see exactly which calls cross the line.
Bifrost syncs model prices every 24 hours by default. To pick up the new rates right now, open the dashboard and hit Force Sync Now.
Supported providers
Both models are live today on the OpenAI API, and Bifrost routes to them with your existing OpenAI keys.
Provider | GPT-6.2 Astra | GPT-6.2 Sol |
|---|---|---|
OpenAI |
|
|
The GPT-6 line also shipped on Microsoft Foundry and Amazon Bedrock after its OpenAI launch. When 6.2 deployments land there, add them to the same virtual key. Bifrost then load-balances across providers and fails over automatically if one rate-limits or goes down.
What changes from GPT-6 Astra and GPT-6.1 Sol
The API surface is the same, so most teams only swap the model string. Check these before moving production traffic:
- Re-tune reasoning effort. OpenAI's headline Astra numbers compare 6.2 at
xhighagainst GPT-6 Astra atmax. Don't carry your old setting over blindly. Start athighand measure. - Answers are more concise. OpenAI links its concise-answer training to more omitted details, and the prompt Codex uses to counter this is not applied to direct API calls. If your app needs exhaustive output, say so in the system prompt and eval for completeness.
- Expect more refusals on security work. Both models launch with tighter safeguards for higher-risk cybersecurity requests. If you run security tooling, watch refusal rates in Bifrost logs after the switch.
- Sol is not a strict upgrade on every task. Artificial Analysis shows Sol three points lower than GPT-6.1 Sol on its Coding Agent Index. Run your own coding evals before a full cutover.
- Long context costs more. Requests over 272K input tokens bill at the long-context tier for the whole request. Set budgets per virtual key so a runaway agent can't surprise you.
Roll out gradually. Put GPT-6 Astra and GPT-6.2 Astra (or 6.1 Sol and 6.2 Sol) under one virtual key with weighted routing, say 90/10. Run your evals, then compare cost, latency and output tokens per model in Bifrost logs before shifting more traffic.
Get started
No Bifrost yet? Start the gateway locally in one command, then add your OpenAI key in the dashboard at http://localhost:8080.
OpenAI SDK (Responses API) through Bifrost's /openai route
OpenAI-compatible Chat Completions
Swap the model string to move between Astra and Sol, or to fall back to any other provider Bifrost supports. Same request shape either way.
Ready to build reliable AI applications?
Run GPT-6.2 Astra and Sol behind one gateway with budgets, guardrails, failover and full observability. Book a demo or star the open-source repo on GitHub.