On 30 July 2026 OpenAI cut the price of GPT-5.6 Luna by 80 percent, to 0.20 dollars per million input tokens and 1.20 dollars per million output. Terra came down 20 percent. Sol’s base price did not move, but it gained a Fast mode at twice the standard rate.

Taken alone that is a discount. Taken alongside what everyone else shipped in the same six weeks, it is the clearest signal yet that the industry has run out of easy ways to differentiate on capability.

The capability spread has compressed

Look at what a buyer sees in August 2026. GPT-5.6 offers 1,050,000 tokens of context. Claude Opus 5 and Sonnet 5 offer 1,000,000. Gemini 3.1 Pro offers 1,048,576. DeepSeek-V4-Flash, Qwen3.8-Max, and Kimi K3 are all in the same range. Every one of them posts strong results on the agentic coding benchmarks that currently matter most.

Two years ago the answer to which model should I use was a capability question with a defensible answer. Today, for a large share of real work, several models will do the job. Once that is true, procurement stops asking which is best and starts asking which is cheapest for an acceptable result.

Three different ways to compete on price

Cut the sticker price. This is what OpenAI did with Luna. It is the most legible move and the easiest for a competitor to match.

Use fewer tokens. xAI reports that Grok 4.5 uses roughly 4.2 times fewer tokens than comparable models on its benchmarks. Google reports that Gemini 3.6 Flash uses 17 percent fewer output tokens than its predecessor on the Artificial Analysis Index. If those figures hold on real workloads, they change the effective cost far more than a headline discount does, because they apply to every request rather than to a rate card.

Make repetition nearly free. Cached input pricing is where the real spread sits. OpenAI charges 0.50 dollars per million cached input against 5.00 fresh, a tenth. DeepSeek charges 0.0028 dollars per million on a cache hit against 0.14 on a miss, a fiftieth. Most production systems send largely the same context repeatedly, so this is the number that decides a monthly bill, and it is the number least often quoted in a comparison.

What this does to the market

Three consequences look reasonably safe to predict.

The middle tier gets crowded. Terra, Sonnet 5, Gemini 3.6 Flash, Grok 4.5, and Qwen3.8-Max are all competing for the same buyer: someone who needs good enough at volume. Anthropic put Sonnet 5 on its free plan. OpenAI made Luna the default for Free and Go users. Free frontier class access is now a distribution strategy rather than a loss leader.

Open weights become a pricing floor. DeepSeek-V4-Flash is MIT licensed with a million token context. Any proprietary vendor pricing a comparable model has to justify the gap. DeepSeek’s own pricing page warns of a significant increase ahead, which suggests even it does not consider the current rate sustainable.

The frontier tier stays expensive, and narrows. Claude Fable 5 sits at 10.00 dollars input and 50.00 output, fifty times Luna’s input rate. That price only survives if the work genuinely requires it. Anthropic’s own numbers show Opus 5 landing within 0.5 percent of Fable 5 on CursorBench 3.2 at half the cost, which is an unusually honest thing for a vendor to publish and a strong hint about where most buyers should land.

What buyers should actually do

Stop comparing price per million tokens. Compare cost per completed task, measured on your own work. A model at three times the rate that gets it right first time is cheaper than one at a third of the rate that needs two retries and a human review.

Then check three things before signing anything: where the long context pricing tier boundary sits, what the cached input rate is, and whether the introductory price you were quoted expires. Anthropic’s Sonnet 5 introductory rate was scheduled to end on 31 August 2026 and rise by 50 percent. That was disclosed at launch, and it is the sort of detail that turns a good deal into a surprise. Anthropic cancelled that increase on 10 August 2026 and made the introductory price permanent, which does not change the advice: read the expiry date before you sign, because a vendor is free to move it in either direction.

Update

22 August 2026: Anthropic cancelled the scheduled Sonnet 5 price rise on 10 August 2026, three weeks before it was due, and made the 2.00 and 10.00 dollar introductory rates permanent. This piece originally reported the increase as scheduled, which was accurate when it was published on 7 August. The passage under “What buyers should actually do” has been amended to reflect what happened. We resolved the corresponding watchlist question in Anthropic cancelled the Claude Sonnet 5 price rise.