The short answer
GPT-6.1 Sol is a price-performance reset: OpenAI claims near-Astra results on coding, computer use and office work at $2 input and $10 output per million tokens, one-fifth of GPT-6 Astra’s standard rates. For teams building on the OpenAI API, the new model ID is gpt-6.1-sol, and it is already live in ChatGPT Work and Codex for paid business plans.
The engineering takeaway is not “switch everything.” It is that many agent workloads now sitting on Astra for quality can probably move down a tier, and the only way to know which ones is to run them side by side on your own tasks and compare cost per completed job.
What did OpenAI ship with GPT-6.1 Sol?
GPT-6.1 Sol is a mid-tier model that OpenAI positions directly against its own flagship. The company describes it as a major upgrade to GPT-6 Sol, with the biggest gains in agentic coding, computer use and multi-step professional work such as debugging, document understanding and workflow execution. CNBC noted the release came just one week after GPT-6 Sol itself, and that CFO Sarah Friar used the stage to highlight the price cut.
Pricing is the headline. Standard API rates are $2 per million input tokens, $0.10 per million cached input tokens and $10 per million output tokens. VentureBeat, which compared the rate cards, puts GPT-6 Astra at $10 and $50 for the same tiers, so uncached input and output cost exactly one-fifth as much, and cached input is half of GPT-6 Sol’s previous $0.20.
The launch came with platform changes that matter for agent builders. OpenAI says Codex can now run on a local machine, remotely from a phone or in the cloud, and TechCrunch reports reusable cloud environments that follow a developer across devices. The Agents API gained computer use, bringing Codex-style multi-agent orchestration, tool search and context compaction into custom applications.
How close is Sol to Astra, really?
On OpenAI’s own evaluations, close. VentureBeat reports that GPT-6.1 Sol beats GPT-6 Sol’s best DeepSWE v1.1 result by 6.4 percentage points, improves AutomationBench by 4.8 points and lands within 2.1 points of Astra on OSWorld 2.0. TechCrunch adds that the factual-error rate on hard prompts drops from 11.4% to 7.7% at low reasoning effort and stays within 1.9 points of Astra across reasoning settings.
These are vendor-run numbers, and the gap that remains tends to show up exactly where teams feel it: long, ambiguous tasks with many tool calls and little supervision. TechCrunch also reports, citing The Wall Street Journal, that a GPT-6.1 Astra update was shelved over safety concerns, so Astra itself is not moving for now. That makes Sol the model where OpenAI’s near-term improvements will land.
Speed is a separate choice. OpenAI says GPT-6.1 Sol Ultrafast will arrive in the coming days with up to 8x faster generation in Codex and 6x in the API. VentureBeat reports up to 300 tokens per second, priced at six times the standard rate. That is useful for interactive pair-programming and live assistants, and wasteful for background agents that nobody is watching.
What it means for US & EU software teams
First, model routing becomes the main cost lever. A fivefold price gap between two models that score within a few points of each other is too large to ignore. The practical pattern is a router: send well-scoped coding, triage, extraction and document work to Sol, and keep Astra for the long-horizon tasks where your evaluations still show a clear quality win.
Second, measure cost per task, not per token. A cheaper model that needs more retries, longer outputs or extra tool calls can erase its discount. Track tokens, tool calls, latency and success rate per completed task, per workflow and per reasoning effort. That is also how you decide whether Ultrafast pays for itself in a specific user-facing flow.
Third, the release cadence is now a governance issue. Two Sol versions in one week means model behavior can shift faster than most change-management processes. Under the EU AI Act’s documentation duties, GDPR accountability and typical US enterprise vendor reviews, you are expected to know which model produced an output. Pin explicit model IDs in production, log the model reported on every response and version your evaluation results alongside your code.
What to do before you switch
- Re-run your eval set. Compare GPT-6 Sol, GPT-6.1 Sol and GPT-6 Astra on your own tasks, including output format and length, before moving traffic.
- Measure cost per completed task. Record tokens, cached-token share, tool calls, retries and latency; that is where the one-fifth price shows up or disappears.
- Route by workflow. Move the workflows that pass to Sol first and keep Astra as a named route for the ones that do not.
- Reserve Ultrafast for latency. Use it only in interactive flows where speed changes user outcomes; keep batch and background agents on standard.
- Keep a rollback. Pin the previous model ID as a fallback route until your own metrics confirm the upgrade.
Frequently asked questions
What did OpenAI release on September 29, 2026?
At DevDay 2026 OpenAI released GPT-6.1 Sol, an upgrade to GPT-6 Sol that shipped one week earlier. OpenAI says it nearly matches its flagship GPT-6 Astra on agentic coding, computer use and professional work. It is available in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users, and in the API under the model ID gpt-6.1-sol.
How much does GPT-6.1 Sol cost in the API?
Standard pricing is $2 per million input tokens, $0.10 per million cached input tokens and $10 per million output tokens. That is one-fifth of GPT-6 Astra's $10 and $50 standard rates, and VentureBeat reports the cached-input price is half of GPT-6 Sol's $0.20. OpenAI says an Ultrafast tier with much faster token generation is coming in the following days at a premium.
Can GPT-6.1 Sol replace GPT-6 Astra for coding agents?
For many workloads it may. VentureBeat reports OpenAI's figures put GPT-6.1 Sol within 2.1 percentage points of Astra on OSWorld 2.0, and TechCrunch reports its factual-error rate stays within 1.9 points of Astra across reasoning settings. These are vendor-run results, so compare both models on your own tasks before moving production traffic.
What is GPT-6.1 Sol Ultrafast?
Ultrafast is a higher-speed serving tier OpenAI says will arrive in the coming days, with up to 8x faster token generation in Codex and 6x in the API. VentureBeat reports speeds of up to 300 tokens per second and pricing at six times standard rates, so it suits latency-critical interactive work rather than batch or background agents.
What should engineering teams do before switching to GPT-6.1 Sol?
Treat it as a model change, not a string swap: re-run your evaluation set against GPT-6 Sol and Astra, measure cost and latency per completed task, route only the workflows that pass, log which model served each request, and keep the previous model pinned as a rollback until your own metrics confirm the gain.
Sources
OpenAI — Introducing GPT-6.1 Sol (company announcement)
OpenAI — DevDay 2026 recap
TechCrunch — OpenAI launches GPT-6.1 Sol, says it nearly matches GPT-6 Astra and costs less
CNBC — OpenAI DevDay 2026: live updates and announcements
VentureBeat — GPT-6.1 Sol offers Astra-like performance at 1/5th price; Ultrafast tier clocks 300 tokens per second