GPT-5.6 Sol gets a three-month API price cut

OpenAI has cut GPT-5.6 Sol API pricing for the next three months, moving the standard short-context rate to $4 per million input tokens and $20 per million output tokens. This is not a new model release, but it is still worth attention because Sol is the expensive tier many teams reserve for hard coding, agentic, research, and multi-step workflow jobs.

The useful engineering takeaway is simple: if you priced Sol out of production agents earlier this summer, rerun the math. The new rate changes the boundary between routing to a cheaper model and paying for the stronger one, especially when the expensive part of your workload is output-heavy code generation, long planning traces, or repeated repair loops.

What was confirmed

OpenAI’s current model documentation lists GPT-5.6 Sol at $4.00 per million input tokens, $0.40 per million cached input tokens, and $20.00 per million output tokens. The same page says the promotional pricing is available at least through November 21, 2026, and compares the new Sol input price with GPT-5.5 at $5.00 per million input tokens.

AWS published a matching Amazon Bedrock announcement on August 21, 2026. AWS says OpenAI lowered GPT-5.6 Sol pricing, and that Sol now costs $4 per million input tokens and $20 per million output tokens. AWS also frames the move as 20% lower input pricing and 33.3% lower output pricing, with promotional pricing available at least through November 21, 2026.

Reuters, republished by Boursorama, reported the same basic numbers on August 21, 2026: previous Sol pricing was $5 per million input tokens and $30 per million output tokens, and the new short-context standard rate is $4 and $20. Reuters also reported that the cuts apply to OpenAI’s API and are rolling out for eligible credit-based usage in ChatGPT Work and Codex, while Pro, Plus, and Business subscription prices remain unchanged.

There are two important footnotes in the OpenAI docs. First, prompts above 272K input tokens are priced at 2x input and 1.5x output for the full request. Second, cache writes are billed at 1.25x the uncached input token rate. If your workload is a large-context agent that drags repository state, tickets, logs, or documentation into every request, do not use the headline price alone.

Why this matters for working engineers

Most teams do not care about model prices in the abstract. They care when a workflow crosses from nice demo to line item.

For coding agents, vulnerability triage, migration assistants, customer-support copilots, and internal research tools, model cost usually hides in a few places:

• repeated retries after failed tool calls;
• large repository or document context;
• verbose chain-of-work style planning output;
• generated patches and tests;
• evaluation runs across many examples;
• background automation that runs whether or not a human is watching.

A one-third cut on output tokens is meaningful because many agent workloads are output-heavy. A model that writes a patch, explains it, updates tests, comments on failures, and then rewrites the patch can burn output tokens faster than a simple Q&A app. If you previously capped Sol to manual escalation only, the new numbers may justify using it for more automated retry or review steps.

That said, cheaper does not mean cheap. Sol is still a frontier-priced model compared with smaller tiers. The right move is not to point every request at Sol. The right move is to revisit your router.

What this changes in practice

Start with your routing policy. If your current policy says use Sol only when Terra or Luna fails twice, the output discount may make one earlier Sol call cheaper than several weaker attempts. This is especially true for tasks where failure is expensive: broken code edits, incorrect database migrations, bad incident summaries, or noisy security findings.

Next, update your evaluation budget. Many teams run small evals because frontier model sweeps are costly. A temporary price cut is a good window to expand test coverage around your real workload: repo-scale edits, tool-calling reliability, JSON schema compliance, code review quality, and regression cases where smaller models struggled.

Then check long-context behavior. OpenAI’s documentation says requests above 272K input tokens get long-context pricing applied to the full request, not just the excess. That creates a cliff. If your agent sometimes crosses that boundary by stuffing in a whole repo map or too many retrieved chunks, you may erase the savings without noticing.

Caching deserves another pass too. The cached input price is listed at $0.40 per million tokens, while cache writes cost more than uncached input. For stable prompts, policy blocks, repository summaries, dependency maps, or API documentation that repeats across calls, caching can be a real lever. For highly variable prompts, it can be noise. Measure cache hit rate before celebrating.

Codex users should also treat this as a quota-model change, not just an API price change. Reuters reported that the cut is rolling out across eligible credit-based plans for ChatGPT Work and Codex, while fixed subscriptions are unchanged. If your team runs Codex through subscription credits rather than direct API billing, watch the credit card for Sol tasks, not only the public API pricing page.

A practical migration checklist

First, snapshot last week’s usage by model, endpoint, input tokens, cached input tokens, output tokens, and request length. Split anything above and below 272K input tokens.

Second, recalculate the same workload at the new Sol price. Do this per workflow, not just globally. The interesting question is not whether the bill drops. The interesting question is which workflows are now cheap enough to run with stronger reasoning.

Third, compare Sol once versus cheaper-model retries. For agentic systems, retries are often the hidden tax. If a cheaper model needs three attempts and still hands work to a human, it may not be cheaper.

Fourth, add a temporary budget guard that expires before November 21, 2026. This is promotional pricing, and both OpenAI and AWS use the phrase at least through that date. Build the rollback into your config now so you are not surprised later.

Fifth, refresh any customer-facing pricing, internal chargeback, or cost-per-ticket assumptions. If you sell AI-assisted workflows, your margin model changed. If you run internal platform engineering, your showback numbers changed.

My read

This is a price-performance story, not a capability story. The model did not become more accurate because the token price changed. What changed is the set of workloads where it is rational to try the stronger model first.

For small teams, the best use of this window is experimentation: broaden evals, test agent routing, and find the tasks where Sol actually reduces total work. For larger teams, this should trigger a billing and policy review. Any hardcoded model selection, static per-task budget, or stale internal price sheet is now suspect.

I would not rebuild an architecture around a promotional price. I would use the next three months to collect evidence. If Sol materially reduces retries, manual review, or failed automation in your own stack, you will have a real case for keeping it in the path even if the promotion changes later.

Sources

• GPT-5.6 Sol model documentation, OpenAI API
• Amazon Bedrock announces reduced pricing for OpenAI GPT-5.6 Sol, AWS
• OpenAI réduit de plus de 20 % les tarifs réservés aux développeurs pour le modèle de pointe GPT-5.6 Sol, Reuters via Boursorama
• OpenAI cuts GPT-5.6 Sol API and credit prices, Techmeme aggregation of Reuters and related coverage
• Advancing the price-performance frontier with GPT-5.6, OpenAI background on July Terra and Luna cuts