At DevDay 2026 in San Francisco, OpenAI announced 20+ updates, and almost all of them serve one idea: make models cheaper and faster, and turn the ChatGPT subscription into a payment layer for an app ecosystem. The main new agent-line item is always-on dots on the GPT-6 Astra model, each with its own cloud computer and plugins for more than 4,000 apps. Following GPT-6 Sol, the company released GPT-6.1 Sol just a week later at a price of $2 per million input tokens and a claimed "nearly Astra level," while the Pro 500 tier unlocks access to the Ultrafast mode with performance up to 300 tokens per second. The independent Artificial Analysis rating places the new model just one point below Astra, so on this release OpenAI's price advantage is better confirmed than its scoring advantage.

image
image

What happened

OpenAI at DevDay 2026 in San Francisco simultaneously presented more than twenty updates. The key product line is dots, always-on agents on the GPT-6 Astra model: each agent gets its own cloud computer and works in 4,000+ apps via plugins. The first dot is included in Pro and Business Premium plans, conversations with it do not consume subscription limits, and in Enterprise special agents are managed through Microsoft Agent 365. At the same time, the GPT-6.1 Sol model with the identifier gpt-6.1-sol was released — just 7 days after GPT-6 Sol, at a price of $2 per million input tokens and $10 per million output tokens; on the Artificial Analysis composite index it scored 52 points versus 53 for Astra and 58 for the rating leader Claude Opus 5.5. The pricing line was also updated: Pro 500 at $500 gives 25x Plus limits and access to Ultrafast mode, which boosts Astra to 300 tokens per second — 8x faster in Codex and 6x in the API, but 6x more expensive, at $300 per million output tokens, while Ultrafast for Sol is promised later. Sign in with ChatGPT already covers 16 apps, including Devin, Notion, Vercel, and OpenClaw, and the Decisions API on the Luna model is open in limited preview with a promise of further rollout.

Context

DevDay 2026 is better understood through three processes that were already underway before the conference. The first is the price race with Anthropic: GPT-6.1 Sol appeared as a direct response to the price cut of Claude Opus 5.5 a week earlier, and its price sets a new lower bound for the flagship class of models. The second is tempo: such a short cycle from release to release technically points to an incremental update like fine-tuning or distillation, rather than a new pretrain, although OpenAI did not disclose the model creation methodology. The third is platform restructuring: OpenAI is turning the ChatGPT subscription into a payment and distribution layer for the ecosystem, modeled on "Sign in with Apple," and is layering its inference architecture — generation remains with large models, while classification, routing, and selection from a fixed set of answers move to cheap specialized endpoints like the Decisions API on the Luna model. In such a picture, specific model versions become swappable, and the durable asset for teams is their own comparison and routing infrastructure on top of composite indices.

Why this matters for the industry

For the industry, three business signals are important here. Inference is rapidly being commoditized: the release and pricing cycle has shrunk to one week, and Sol delivers approximately Astra level on the composite index for a fifth of the price — token budgets, contracts, and unit-economics assumptions will now have to be constantly revised. Sign in with ChatGPT turns the subscription into a payment layer for the ecosystem: the developer no longer pays for tokens directly, and the dots plugin store gets access to a base of 1.2 billion weekly ChatGPT users. The Decisions API moves classification and routing out of expensive LLM generation into a cheap endpoint with a fixed set of answers — this is direct price and speed pressure on specialized startups like Jev, and if the endpoint is rolled out publicly, classifiers and routers risk becoming a commodity layer. Over the long term, what will surface in such processes is not individual "top models" but layers: model selection and routing tools, agent platforms, and evaluation methodologies, where the durable asset becomes one's own evals on top of composite indices.

Why this matters for users

The practical takeaway for engineering teams: model selection becomes per-task. Sol is already available via API at $2 per million input tokens and $10 per million output tokens, so it makes sense to immediately recalculate budgets and run a pilot migration of quality-non-critical workloads, running your own eval set and recording the gap with Astra on your own data, rather than relying on a single composite index. For interactive scenarios, Ultrafast is already working in the API and Codex: up to 300 tokens per second, but with a sixfold price for output tokens, meaning it is a direct tradeoff between speed and budget. End users get sign-in to 16 apps with Sign in with ChatGPT, including Devin, Notion, Vercel, and OpenClaw, under a single ChatGPT account, where subscription limits are consumed; meanwhile, Pro and Business Premium subscribers get the first dot at no extra cost, and correspondence with it does not consume the chat quota at all. For those hitting limits, the Pro 500 tier is addressed: for $500 it gives 25x Plus limits plus access to Ultrafast.

What is still unknown / limitations

The main caveats concern OpenAI's phrasing "nearly Astra level." A one-point gap within the Artificial Analysis composite index (52 versus 53) with undisclosed weights, no confidence intervals, and no per-task breakdown cannot be interpreted as parity — for now this is a manufacturer-stated estimate relying on a single independent source, and the actual gap of the leader Claude Opus 5.5 with a score of 58 from both OpenAI models is about six points, so Sol's price advantage is more confirmed than its scoring advantage. The methodology for creating GPT-6.1 Sol has not been disclosed, so the version about fine-tuning or distillation in a week remains an interpretation, not a fact. The Ultrafast mode has neither its acceleration mechanism nor the definition of the "8x faster" metric disclosed — it is unknown what exactly was measured, time to end of generation or end-to-end latency, and the impact on answer quality has not been shown. The Decisions API is available only in limited preview, public rollout remains a promise, and Ultrafast for Sol is only announced.

Sources

Author

Look at AI, editorial team