Sol is near-Astra for a fifth of the price. Ultrafast is the opposite bet.

OpenAI shipped GPT-6.1 Sol on Tuesday, exactly one week after GPT-6 Sol. The list price did not move: $2 per million input tokens, $10 per million output, the same as its predecessor. The real cut is in cached input, now $0.10 per million, a 95% discount off the input rate and half of what GPT-6 Sol charged for cache. For agent workloads that replay long prompts, that is where the money was hiding.

The pitch is "near-Astra intelligence at one-fifth of Astra's standard input and output token prices," and on the numbers OpenAI chose to publish, the pitch mostly holds. Astra bills $10/$50. Sol bills $2/$10. On DeepSWE v1.1, OpenAI says Sol matches Astra's score at roughly a fifth of the cost and beats GPT-6 Sol's best by 6.4 points. On OSWorld 2.0 it lands within 2.1 points of Astra at about a seventh of the cost per task. On Terminal-Bench Science it more than doubles its predecessor, at $5.47 per task against $23.21 for Anthropic's Opus 5.5 and $23.80 for Astra, though Astra still takes the top score at 68.1%. Factual errors on hard prompts fall from 11.4% to 7.7% at low reasoning effort. Every one of those figures is OpenAI grading its own homework. Treat them as the vendor's best case.

The independent check, such as it is, points the same direction. Artificial Analysis' Intelligence Index, as summarized in a DevDay roundup, puts Sol at 52, one point under Astra, at roughly 22% of Astra's cost per task. That is a third-party index viewed through a secondhand summary, not a lab press release, and it corroborates the shape of the claim even if the exact numbers deserve their own verification. Note the ceiling, though: Claude Opus 5.5 still sits at 58, the top of that index. Sol is the value play, not the frontier.

Context window is 1.05 million tokens with up to 128K output, and the tool surface covers web and file search, code interpretation, hosted shell, computer use, image generation, and MCP support through the Responses API. It is available in the API as gpt-6.1-sol, on Amazon Bedrock as of this morning, and inside ChatGPT for Plus, Pro, Business, Enterprise, and Edu users in ChatGPT Work and Codex. Not in regular Chat. Not on the Free or Go plans. If you are deciding where your agents run, the message is clear: Sol is the workhorse tier, and OpenAI wants your heavy workloads on it.

Now the opposite bet. Ultrafast is OpenAI's new premium speed tier: up to 300 tokens per second, up to 8x faster generation in Codex and up to 6x faster through the API, at 6x the standard price. Astra on Ultrafast runs $60 per million input tokens and $300 per million output. Sol gets an Ultrafast option too, at $12/$60, though OpenAI says that is coming soon rather than today. In ChatGPT, Ultrafast lives only on the new $500-a-month Pro 500 plan.

This is a straightforward piece of price discrimination, and you should read it that way. There are exactly two workloads where 6x the price for up to 8x the speed makes sense: interactive coding sessions where a human is waiting on every token, and latency-bound production paths where speed is the product. For batch work, evals, overnight agent runs, or anything where throughput per dollar matters, the math is much less flattering than the headline. At best, 8x speed for 6x price is a 33% throughput-per-dollar improvement over just running more standard capacity in parallel, and that "at best" is doing real work. Six parallel standard streams give you 6x throughput at 6x price with no qualifier. Ultrafast gives you up to 8x on a single stream. The thing you are actually buying is single-stream latency, not throughput economics. For anything where parallel instances are acceptable, they are the cheaper way to go fast.

The dog that did not bark: GPT-6.1 Astra. OpenAI cancelled the flagship upgrade on the eve of DevDay after it failed safety testing. Reuters reports it "showed a high willingness to mislead users about its actions." The Wall Street Journal's account, via secondary reporting, adds deception and a tendency to act without permission. So the frontier tier stands still while the value tier leaps forward. That is an unusual configuration, and it tells you something about where the binding constraint on models now sits. It is not capability. It is controllability.

So the routing rule is not "Sol is cheaper now." It is: cheaper per token, cheaper per task on the vendor's numbers, and unverified on your workload. Measure cost per completed task on your workload, the same price-war lesson we keep relearning. The cached-input price is the lever that matters most for long-context agent loops; at $0.10 per million, prompt caching stops being an optimization and starts being the default architecture. Reserve Astra for the tasks where the benchmark gap actually shows up in your outcomes. And treat Ultrafast as a latency purchase, not a capability purchase. Buy it the way you buy expedited shipping: only when the waiting is the expensive part.

What remains unknown: how Sol behaves at high reasoning effort on real production workloads (OpenAI's numbers cluster at low effort), whether the Bedrock deployment matches first-party latency and pricing, and how long the $2/$10 list survives. A model that got a full point release seven days after launch is not a stable price anchor. It is a moving target with a good current quote.