GPT-5.6 Sol API Price Falls to $4 Per Million Tokens

Frontier model pricing moved again this week, and this time it moved down. OpenAI cut developer pricing for GPT-5.6 Sol on August 21, 2026, taking more than 20% off input tokens and a third off output tokens. The reduction runs as a promotional rate through November 21, 2026, and applies to metered API usage rather than to what subscribers already get bundled in their plans.

• Input tokens fall from $5.00 to $4.00 per million, a 20% reduction
• Output tokens fall from $30.00 to $20.00 per million, a 33% reduction
• Cached input reads drop from $0.50 to $0.40, and cache writes from $6.25 to $5.00 per million
• The promotional pricing is scheduled to run through November 21, 2026

What the New GPT-5.6 Sol Pricing Actually Changes

The standard rates above apply to requests up to 272K input tokens. Above that threshold, the long-context tier still charges roughly double, but it takes the same percentage reductions, so the discount follows you into large-context work rather than evaporating at the boundary.

The asymmetry between the two headline numbers is the part worth studying. Input fell 20%, output fell 33%. For a summarization or retrieval workload that reads a great deal and writes a little, the effective saving lands closer to 20%. For an agentic coding loop that generates long diffs, test files, and explanations, the saving lands closer to a third. OpenAI has effectively made the generation-heavy half of the market cheaper, which is where reasoning models burn the most tokens.

Caching gets the same treatment. Cached reads at $0.40 per million make repeated context, such as a large system prompt or a fixed codebase preamble, roughly a tenth the price of fresh input. Any team running a long-lived agent should already be structuring prompts to hit that cache, and the new rate widens the gap further.

Who Gets the Discount

The reduction applies to API usage and to credits on eligible plans covering OpenAI's agentic product and its coding tool. Included subscription usage on Pro, Plus, and Business accounts is unchanged, which is a sensible distinction: those tiers are priced on capacity, not per token.

Reuters framed the move as a response to competitive pressure from both Anthropic and a wave of capable Chinese open-weight models. That reading fits the pattern. This is the second pricing adjustment in the GPT-5.6 family since it launched, following earlier reductions we covered when OpenAI cut Luna prices after rewriting its GPU kernels at the start of August. If you are trying to work out which member of the family fits your workload, our breakdown of the Sol, Terra, and Luna tiers still maps the trade-offs.

What Should Developers Do With Three Months?

Treat it as a window rather than a new baseline. The rate is scheduled to expire on November 21, 2026, and OpenAI has not said what happens after that.

The practical move is to use the window for the work that was previously too expensive to justify. Batch evaluations, larger retrieval contexts, and longer agent runs all become defensible at a third off output. Benchmark those workloads now, while the pricing is favourable, so you have real numbers on what they cost and what they deliver before you commit to them at whatever rate follows.

It is also worth measuring rather than assuming. A 33% cut on output only reaches your invoice if output is where your tokens go. Pull a week of actual usage, split it by input, cached input, and output, and apply the new rates. Teams are regularly surprised by how much of their bill is uncached system prompt.

For more on where frontier pricing is heading, see our ongoing AI and machine learning coverage.

Sources: Reuters via Investing.com — August 21, 2026; WinBuzzer — August 23, 2026.