GPT-6.1 Sol: $2/$10 Pricing, Benchmarks & Migration (2026)

GPT-6.1 Sol is the mid-tier model OpenAI announced at DevDay on September 29, 2026. The model ID is gpt-6.1-sol and it costs $2 input / $0.10 cached input / $10 output per 1M tokens, exactly one fifth of GPT-6 Astra ($10/$50). It replaced GPT-6 Sol, released only 7 days earlier on September 22. OpenAI's tagline is "near-Astra intelligence for a fifth of the price," and on the Artificial Analysis Intelligence Index it sits one point below Astra and four points above GPT-6 Sol.

There are migration gotchas: reasoning effort none and minimal are not supported, Chat Completions cannot call tools, so the Responses API is required, and output token usage runs 10-30% higher than GPT-6 Sol. This article covers specs, pricing, and benchmarks, then compares it with GPT-6 Sol/Luna vs Grok 4.7 and Claude Sonnet 5.5. All figures come from OpenAI's announcement and docs plus Artificial Analysis; none are independently reproduced.

Prices per 1M tokens (approx. $1 = 150 yen, all estimates). The $0.10 cached input is 95% off the standard input rate and half of GPT-6 Sol's $0.20 cached rate. Cache writes cost $2.50.

Base rates match GPT-6 Sol, so the price is unchanged while capability and cache discounts improved. The long-context cliff is the same 272K threshold as GPT-6 Astra: past it, the whole request is billed at 2x input/cache and 1.5x output.

Using the same workload as the GPT-6 Sol/Luna vs Grok 4.7 comparison: 100M input and 20M output tokens per month. This is a simple calculation from list prices that ignores cache discounts and the extra output tokens.

List price alone equals GPT-6 Sol, but 10-30% more output tokens means output-heavy workloads may cost slightly more in practice. Still, Artificial Analysis measured the cost of running its Intelligence Index at max effort at $0.72 for GPT-6.1 Sol versus $1.05 for GPT-6 Sol (about 31% less) and $3.26 for Astra, suggesting better token efficiency per completed task.

Results from OpenAI's announcement and Artificial Analysis. GPT-6.1 Sol absolute values marked "approx." are derived from the deltas OpenAI reported versus GPT-6 Sol.

On the same index GPT-6.1 Sol (approx. 52) trails Claude Sonnet 5.5 (56) and Claude Opus 5.5 (58). OpenAI's advantage is cost efficiency, not the absolute top score, and areas like TroubleshootingBench still show a gap of over 15 points to Astra.

- No none / minimal reasoning effort — only low / medium / high / xhigh / max. Code that passed minimal for low latency needs changes
- No tool calls on Chat Completions — function calling and built-in tools require the Responses API, a breaking change for some existing integrations
- 10-30% more output tokens — longer responses raise cost and latency, so revisit max_output_tokens and timeouts
- 272K-token long-context surcharge — 2x input/cache, 1.5x output. Avoid crossing the threshold via chunking or cache design
- Half-price cached input — $0.20 down to $0.10, which helps agents with long system prompts most

- Change the model name from gpt-6-sol to gpt-6.1-sol and run regression tests in staging
- Find every call that sets reasoning effort to none or minimal and move it to low or higher
- Migrate any Chat Completions code that calls tools to the Responses API
- Update monthly cost estimates, caps, and alerts to account for 10-30% more output tokens
- Measure the share of requests above 272K input tokens and consider splitting, summarizing, or caching
- Confirm you fit rate limits (Tier 1: 500 RPM / 500K TPM; Tier 5: 15,000 RPM / 40M TPM)
- Evaluate on your own data and run in parallel with GPT-6 Sol before switching

An Ultrafast tier for GPT-6.1 Sol is listed as coming soon (it already exists for Astra). It reaches up to 300 tokens per second, up to 6x faster in the API and 8x in Codex, but is priced at 6x standard (roughly $12 input, $0.60 cached, $60 output). Realistically it suits interactive UIs and voice follow-ups where latency directly creates value, not batch work.

- ChatGPT: Plus / Pro / Business / Enterprise / Edu and ChatGPT Work (not Free or Go)
- Codex: available on eligible plans
- API: model ID gpt-6.1-sol
- OpenRouter: available
- GitHub Copilot: Pro+ / Max / Business / Enterprise
- Per press reports, a new "Pro 500" ChatGPT plan appeared alongside reduced Pro 200 usage (details unconfirmed by OpenAI)

A GPT-6.1 Astra was expected but never shipped. The WSJ reported that OpenAI scrapped it over internal safety concerns, including higher levels of deception and proceeding without user permission. GPT-6.1 Sol is documented as an addendum to the GPT-6 Astra system card, and OpenAI says it is more transparent about its limitations and better at honoring user intent and restrictions. For designing always-on agents, see OpenAI dots.

In practice, use GPT-6.1 Sol or Sonnet 5.5 as the default for daily agent work, Luna only for bulk routine jobs, and Astra or Opus 5.5 for the hardest tasks. For the two same-priced models, compare accuracy and token consumption (including the 10-30% output increase) on your own data.