GPT-6.1 Sol Guide: OpenAI's Middle Tier at a Fifth of Astra's Price

OpenAI's current line-up has three general-purpose models: GPT-6 Astra at the top, GPT-6 Luna at the bottom, and GPT-6.1 Sol in the middle. OpenAI's own one-line description of Sol is "near-Astra performance for complex work at a lower cost", and its GPT-6 guide goes further than vendors usually do: it tells you to compare Sol with Astra on your own tasks.

That is the right instinct, and it frames this guide. Sol is not a cut-down model for easy work. It is a bet that most complex tasks do not need the flagship, priced so that being wrong about a few of them is cheap.

Sol's job in the GPT-6 line-up

OpenAI positions the three models as balanced intelligence and cost for Sol, maximum capability for Astra, and focused high-volume work for Luna. The list prices per million tokens make the gaps concrete:

• GPT-6 Astra — $10 input, $1 cached, $50 output
• GPT-6.1 Sol — $2 input, $0.10 cached, $10 output

Sol is a fifth of Astra's price on both input and output, and its cached-input rate is a tenth of Astra's. The limits are otherwise identical: a 1,050,000-token context window, a 922,000-token input cap, 128,000 tokens of output, text and image in, text out, and a knowledge cutoff of 30 April 2026.

OpenAI names the target work as complex coding, computer use and professional tasks. The Responses API tools available are the same broad set as Astra's, including hosted shell, apply patch, computer use, code interpreter, web search, file search and MCP.

Migrating from GPT-6 Sol: two traps

The 6.1 in the name is a point release over GPT-6 Sol, and OpenAI asks existing gpt-6-sol users to read the migration guidance before switching. Two differences break working code.

No none or minimal effort. GPT-6.1 Sol accepts low, medium (the default), high, xhigh and max. GPT-6 Sol supported none; 6.1 does not. If you used none for latency-critical paths, the guidance is to start at low and compare results.

Tool calling needs the Responses API. Sol is available on Chat Completions, but without tool calling. An agent built on Chat Completions with function tools will not work after a model-name swap; it has to move to Responses first.

Two smaller items from the same guidance: when effort is anything other than none, remove temperature, top_p and the logprobs options; and if you are coming from GPT-5.5 or earlier, prompt_cache_retention is replaced by prompt_cache_options.ttl.

The 272K price cliff

The million-token window has a pricing catch that is easy to miss. For prompts over 272,000 input tokens, OpenAI charges twice the input and cache rates and one and a half times the output rate, applied to the whole request, not just the tokens above the line.

In practice, a 270K-token prompt and a 280K-token prompt differ in cost by far more than 10K tokens' worth. Sol's input rate effectively becomes $4 per million and output $15 once you cross the line. If your workload sits near that boundary, trimming a few thousand tokens of context to stay under it is one of the better optimisations available. Context window cost tradeoffs covers how to decide what to cut.

Cache writes cost 1.25 times the uncached input rate and cached reads 5% of it, so caching pays back as soon as a prefix is reused once. For long agent sessions that re-send the same system prompt and tool definitions on every turn, it pays quickly.

Speed tiers and discounts

Sol has more pricing tiers than earlier models, and they matter for capacity planning:

• Batch and Flex — half the standard rate, for work that can wait
• Fast mode — twice the standard rate
• Ultrafast mode — six times the standard rate, requested with service_tier: "ultrafast" in the Responses API
• Regional processing — a 10% premium where available. OpenAI's guide lists EU data residency for Sol in Fast mode, and US and EU residency in Ultrafast

Ultrafast Sol costs more per token than standard Astra. If you are tempted by it, check whether the latency you need could be met by a smaller model instead.

Sol or Astra: test it, do not guess

The useful question is not which model is smarter. Astra is the flagship; on hard enough tasks it will win. The question is how often your tasks are hard enough for that to show up.

1. Pick twenty real tasks from recent work, weighted towards the difficult ones.
2. Run each on Sol at medium and high, and on GPT-6 Astra at its default.
3. Score pass or fail, and record total tokens including reasoning.
4. Compute cost per passed task for each configuration.

If Sol at high passes nearly as often as Astra, default to Sol and escalate failures. If the gap is wide on exactly the tasks you care about most, route those to Astra and keep Sol for the rest. Before reaching for Astra, also try pro reasoning mode on Sol, which spends more work on the same model; GPT-6.1 Sol Pro covers that option. How to benchmark LLMs on your own work has the harness setup.