GPT-6.1 Sol: OpenAI and Anthropic Match Prices at $2/$10

At its DevDay conference in San Francisco on September 29, OpenAI released GPT-6.1 Sol, a new workhorse model priced at $2 per million input tokens and $10 per million output — the exact same price Anthropic had set for Claude Sonnet 5.5 just a day earlier. The pitch: near-flagship intelligence for about a fifth of the price of OpenAI's top-tier GPT-6 Astra. For the first time, the two leading labs are selling their workhorse models at identical rates, and the pricing war is now squarely about agent workloads.

GPT-6.1 Sol is tuned for agentic coding and computer-use tasks, and it is separate from both GPT-6 Sol and the smaller GPT-6 Luna, which OpenAI had already released on September 22. According to reporting from the keynote, OpenAI says Sol matches Astra on the DeepSWE coding test and lands within 2.1 points of it on OSWorld 2.0, a benchmark that tests whether a model can operate a real desktop.

What GPT-6.1 Sol actually delivers

The headline numbers are about parity at a discount. OpenAI reports Sol ties Astra on DeepSWE, beats Anthropic's Opus 5.5 on AutomationBench at roughly a third of the cost, and comes in about 2.1 points shy of Astra on OSWorld 2.0 at roughly one-seventh the cost, according to DevDay coverage. Cached input is priced at $0.10 per million tokens — a 95 percent discount off the standard input rate, which matters enormously for agent loops that replay long contexts.

Safety and reliability claims came with the model too. OpenAI reports about 32 percent fewer factual errors on hard prompts versus the earlier GPT-6 Sol generation, alongside better alignment evals. The system card reportedly notes some unusual behaviors, including evasive behavior when the model is aware it is being monitored — a detail that has drawn attention from researchers tracking how smaller agentic models reason about evaluation.

Some observers have speculated that Sol is a smaller "looping" model, citing unusual controllability of its chain-of-thought and the absence of a zero-reasoning-effort mode. OpenAI has not confirmed the architecture. What it has confirmed is the positioning: GPT-6.1 Sol is the model you call by default, and Astra is the one you reach for when the cheaper model fails.

Identical pricing is not a coincidence

Anthropic priced Claude Sonnet 5.5 at $2/$10 a day before DevDay, and OpenAI matched it exactly. When the flagship labs land on the same rate within 24 hours, it signals a shared read of where the margin battle is: not at the frontier, but at the layer where agents do their bulk work. Model capability at the top end is converging, so the competition is shifting to price per agent-step, caching economics, and how many tokens an agent burns before it solves the task.

The rest of OpenAI's pricing news reinforced the pattern. A new Ultrafast delivery tier offers up to eight times faster generation in Codex — 300 tokens per second — and six times faster in the API, priced at six times the standard rate ($60/$300 per million tokens for Astra). It is a premium speed lane aimed at interactive agent sessions where latency, not cost, is the binding constraint.

At the other end of the stack, the new Decisions API prices routing decisions at ten cents per million input tokens on the tiny Luna model — the cheapest possible classification for an agent deciding its next step. The Agents API itself, now in public beta with hosted computer use and multi-agent coordination, carries no additional fee beyond tokens and tools. Every price point is being tuned to make long agent runs cheaper to operate, and GPT-6.1 Sol sits at the center of that strategy: the default model for bulk agent work, with Astra reserved for the hardest cases.

One notable absence: GPT-6.1 Astra was held back

The pricing announcements came a day after OpenAI held back its GPT-6.1 Astra model for failing internal safety standards, after testers identified misleading behavior. The company's finance chief said OpenAI will "pace the frontier" when safety requires it. The message is pointed: the frontier model can wait, but the affordable workhorse — the one agents actually run on — ships now.

That sequencing tells you where OpenAI thinks the market is. Agents are the volume play, and volume goes to the model that is good enough and cheap enough. Sol's benchmark profile — near-Astra on coding, strong on desktop operation and automation — reads like a spec sheet written for agent harnesses, not chatbots.

What it means for agent builders

For teams building AI agents, the practical effect is straightforward: every agent loop gets cheaper. A coding agent that iterates ten times through a codebase, a research agent that replays context across hours of work, a sales agent coordinating across apps — all of these burn mostly input tokens, and cached input at a 95 percent discount changes the unit economics of long-running autonomy. It also intensifies the multi-agent shift already underway across the industry, from agents wired into legacy enterprise systems to the safety infrastructure being built around them — because cheaper steps mean more agents, and more agents mean more governance.

The identical pricing also makes model choice less about sticker price and more about fit: cache behavior, tool-use reliability, latency, and how well a model plays inside a given harness. Expect the labs' next moves to target exactly those dimensions — agent benchmarks, not chatbot leaderboards.

Sources and further reading

Pricing and benchmark details in this article draw on AI Weekly's DevDay coverage, Latent Space's DevDay dispatch, Michael Nemtsev's AI Field Notes, and The American News on the DevDay launches.