SAT, AUGUST 22, 2026
Independent · In‑Depth · Practitioner‑Tested
✎ Large Language Models

One Frontier Lab Cut Prices 20 Percent This Week. The Other Raises Them in Nine Days

OpenAI dropped GPT-5.6 Sol from 5 and 30 dollars per million to 4 and 20 on Friday, citing competition from Anthropic and Chinese labs — but only for three months, reverting around 21 November. Claude Sonnet 5 moves the other way on 1 September, from 2 and 10 to 3 and 15, with a tokenizer change that pushes the effective rise on code past the sticker figure.

By AIToolsRecap August 22, 2026 8 min read 38 views
Home Articles Large Language Models ChatGPT OpenAI Cut GPT-5.6 Sol to $4/$20 for Three Mont...
THE VERDICT

● GPT-5.6 Sol is now $4 / $20 per million, down from $5 / $30. Announced Friday 21 August.

● The catch: it lasts three months. Around 21 November it reverts unless OpenAI extends it.

● Claude Sonnet 5 goes the other way on 1 September — $2 / $10 becomes $3 / $15, plus a tokenizer change adding 10 to 35 percent tokens on code.

● Do not migrate on this alone. A three-month promo is not a reason to rewrite a production integration.

What actually changed

OpenAI said on Friday it is cutting developer pricing on its frontier model by more than 20 percent for the next three months, citing competition from Anthropic and Chinese labs.

Model Input / 1M Output / 1M Note
GPT-5.6 Sol $4 (was $5) $20 (was $30) 3 months only
GPT-5.6 Terra $2 $12 Cut 20% on 30 July
GPT-5.6 Luna $0.20 $1.20 Cut 80% on 30 July
Claude Sonnet 5 $2 → $3 $10 → $15 Rises 1 September
Claude Opus 5 $5 $25 Unchanged
Claude Fable 5 $10 $50 Unchanged
Grok 4.6 $2 $6 Doubles above 200K input

The cut applies to the API and rolls out as credits across eligible ChatGPT Work and Codex plans. Pro, Plus and Business subscription pricing is unchanged. That split is deliberate — OpenAI is protecting consumer revenue while discounting for the usage-heavy developers most likely to test another stack.

Do not confuse this with last week's OpenRouter promo

TWO DIFFERENT THINGS, FOUR DAYS APART

On 18 August a Hacker News thread ran with "GPT-5.6 Sol Pricing Cut by 50%". That was an OpenRouter promo on non-BYOK traffic at an effective $2.50 / $15 — not an OpenAI change. OpenAI's own rate was untouched at the time.

Friday's announcement is the real thing: OpenAI's own rate card, $4 / $20, temporary.

The three-month clause is the whole story

A permanent cut and a three-month promotion produce identical spreadsheets and completely different decisions.

If you migrate a production workload to Sol on the strength of $4 / $20, you are betting the price holds past roughly 21 November. Nothing has been said about what happens then. The July cuts to Luna and Terra were framed as permanent, passing on serving efficiency gains. This one is explicitly for three months and explicitly attributed to competition.

That distinction tells you what it is: a customer-retention move aimed at developers currently evaluating Anthropic and the Chinese labs, not a change in what inference costs OpenAI to serve.

What the two directions actually mean

OpenAI cutting while Anthropic raises is not one company winning. It is two companies with different problems.

OpenAI is defending volume. It filed a confidential S-1 in June, its CFO committed to a 2027 listing this week, and reported figures put it near 2 billion dollars a month in revenue while losing roughly 1.22 dollars for every dollar earned. Developer count and usage growth are the metrics that matter in that story.

Anthropic is defending margin. It reported first operating profit of 559 million dollars on 10.9 billion in Q2 revenue, two years ahead of its own schedule. A company that has just proven it can be profitable raises prices; a company that needs a listing narrative buys usage.

What to actually do

If you are... Do this
Already on Sol Nothing. Enjoy three months, diarise 21 November
On Sonnet 5, watching 1 September Model the tokenizer change too. On code the effective rise exceeds the 50% sticker
Tempted to migrate to Sol Price it at $5 / $30, the rate after November. If it still wins, migrate
Running high-volume cheap tasks Luna at $0.20 / $1.20 is a permanent cut and an order of magnitude below all of this
On a consumer subscription Unaffected. Pro, Plus and Business are unchanged
Undecided Route through an aggregator so the next repricing is a config change

The general rule this fortnight keeps demonstrating: price a promotional rate at its post-promotional value. Migration costs engineering time, and three months of savings rarely covers it.

FAQ

What is GPT-5.6 Sol's new price?

4 dollars per million input tokens and 20 dollars per million output, for standard short-context use, down from 5 and 30. Announced Friday 21 August 2026.

How long does the cut last?

Three months, per OpenAI. That puts the revert around 21 November 2026 unless extended. No commitment beyond that has been made.

Does it affect ChatGPT Plus or Pro?

No. The cut applies to the API and to credits on eligible ChatGPT Work and Codex plans. Pro, Plus and Business subscription pricing is unchanged.

Is Claude now more expensive than GPT-5.6 Sol?

Depends which Claude. Opus 5 at 5 and 25 sits above Sol's promotional 4 and 20 but below its standard 5 and 30. Sonnet 5 stays cheaper even after rising to 3 and 15 on 1 September. Fable 5 at 10 and 50 is a different tier entirely.

Was there not a 50 percent Sol cut last week?

No. That was an OpenRouter promotion on non-BYOK traffic at an effective 2.50 and 15, widely misreported as an OpenAI price change. OpenAI's own rate was unchanged until Friday.

Should I switch from Claude to GPT-5.6 Sol?

Not on price alone, and not on a three-month price. Run your own eval on tasks you care about, then price the winner at post-promotional rates. A cheaper model that needs two attempts is not cheaper.

Tags
OpenAIChatGPTGPT-5.6 SolAnthropicClaude Sonnet 5Claude Opus 5GrokAPI PricingCodex2026

Spot an inaccuracy?

We verify facts before publishing and correct errors promptly. If something in this article is wrong or outdated, let us know.

Report an error →