GPT-6.1 Sol Launches at DevDay as OpenAI Holds Back Astra 6.1

OpenAI's DevDay 2026 centered on a mid-tier model rather than a new flagship. On September 29, 2026, the company shipped GPT-6.1 Sol. It is an update to GPT-6 Sol that OpenAI says comes close to GPT-6 Astra on agentic coding, computer use and professional tasks, at one-fifth of Astra's standard token prices.13 The release came only a week after GPT-6 Sol itself.2 Coverage also points to a notable gap: a 6.1 version of Astra was apparently held back.1

What shipped

GPT-6.1 Sol is available in the API as gpt-6.1-sol. It is also live in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu users.13 It keeps GPT-6 Sol's pricing of $2 per million input tokens and $10 per million output tokens. The cached-input rate drops from $0.20 to $0.10 per million.12 GPT-6 Astra lists at $10 input and $50 output, with cached reads at $1.00.12

One set of figures puts a typical heavy agent turn at about $0.60 on GPT-6.1 Sol. That turn is 200K tokens in and 20K out. With 90% of the input cached, the cost falls to roughly $0.26. The same turn on Astra costs $3.00, or $1.38 cached.1 The base math holds up. At list prices, 200K input tokens cost $0.40 and 20K output tokens cost $0.20.

The competitive comparison matters as much as the internal one. GPT-6.1 Sol matches Claude Sonnet 5.5 on input and output price. Its cached reads are half the price of Sonnet's, and it costs half as much as Claude Opus 5.5 on input and output.1 For developers running long, context-heavy agent loops, cache-read pricing often matters more than headline rates, and that is where this release improves things most.

The catch: where you can use it

One outlet highlights a limitation in its framing: near-Astra power, "but not in chat."2 The listed availability is ChatGPT Work, Codex and the API.13 That suggests OpenAI is aiming this model at professional and developer workflows rather than the general consumer chat experience. Developers on the API won't notice the restriction. Users hoping the cheaper, stronger model would show up in their everyday chat window may be disappointed.

The rest of DevDay

The model launch sat alongside a broad push into agent tooling.3 The Agents API now supports computer use, so applications can operate software through graphical interfaces. The API also absorbs Codex's multi-agent capabilities, along with tool search, tool calling and context compaction. OpenAI runs the execution infrastructure underneath.3

Codex got several updates:3

• Cloud environments: Codex can now run remotely, not just on local machines, so tasks can be started from other devices.
• CLI upgrades: the CLI adds voice input and an /agents interface for delegating and monitoring multiple tasks.
• Code review: a new workflow analyzes diffs and flags issues in GitHub pull requests and GitLab merge requests.
• Codex Security Cloud: it scans repositories and new commits, investigates findings, removes duplicates and prepares fixes.

OpenAI also announced a Decisions API and new plugin capabilities for ChatGPT.3

Why Astra 6.1 is missing

The missing piece is a 6.1 Astra. One release guide includes a section on why the Astra 6.1 update was held back, citing outside reporting.1 The available detail on the reasons is thin, and OpenAI's own messaging focused on Sol. Any account of specific safety failures should be treated as reported rather than confirmed until the company says more.

The shape of the launch still tells a story. OpenAI kept the Sol price flat, cut caching costs and marketed the model by how close it gets to Astra. That positions GPT-6.1 Sol as the default workhorse for agentic development, with Astra as an expensive ceiling few teams will need to reach.123 Whether that was the plan all along or a pivot after the flagship update stalled, the outcome for developers is similar: the most attractive new model at DevDay was the cheap one.

The takeaway

GPT-6.1 Sol appears to be the practical release of this cycle. Its pricing undercuts Anthropic's comparable tiers on cached workloads, it ships into the tools developers actually use, and it pairs with agent infrastructure built for long-running, multi-step tasks. Its "near-Astra" claims rest on OpenAI's own evaluations, so independent benchmarking is the next thing to watch. The other open question is whether, and in what form, a 6.1 Astra eventually ships.