August 22, 2026 ·

openaigpt-5-6pricinganthropicapiprocurementcoding-agentsprice-war

OpenAI cut GPT-5.6 Sol to $4/$20 — and put a 21 November expiry date on the frontier price crown

TL;DR: On 21 August 2026 OpenAI cut developer pricing for its frontier model GPT-5.6 Sol by more than 20%: input $5 → $4 per million tokens, cached input $0.50 → $0.40, output $30 → $20. Reuters confirms the rate covers the API, Codex credits and eligible ChatGPT Work plans, while Pro, Plus and Business subscriptions are unchanged. Sol now undercuts Claude Opus 5 ($5/$25) on both input and output for the first time. But OpenAI’s own announcement dates the rate 21 August to 21 November 2026 and says nothing about day 92. That expiry is the story. Three of the five prices buyers quote most often are now promotional — Sol to 21 November, Gemini 3.7 Flash to 31 December, DeepSeek’s cheap rate only outside peak hours. Eight days ago our read of this market named Opus 5 the frontier value pick. It is no longer true, and it took eight days to stop being true. The frontier price leaderboard now turns over faster than a procurement cycle closes — which means the right response to this cut is not to re-route to Sol, it is to stop treating a dated discount as a fact about your cost base.

What changed

OpenAI announced on Friday that it is “dropping API and credit pricing of GPT-5.6 Sol by over 20% for the next 3 months.”

GPT-5.6 Sol, per 1M tokens	Was	Now	Change
Input (standard context)	$5.00	$4.00	−20%
Cached input	$0.50	$0.40	−20%
Output	$30.00	$20.00	−33%

The asymmetry is the first thing worth noticing. Output falls a third while input falls a fifth, and for agentic workloads — long reasoning traces, multi-step tool calls, code generation — output is where the money goes. On a typical coding-agent mix, the effective saving lands closer to the 33% than the 20%.

Scope. Reuters reports the cut is effective on the API and is rolling out across eligible plans for credits on ChatGPT Work and Codex. Consumer and business subscription pricing — Pro, Plus, Business — is unchanged. If your team consumes Sol through seats rather than through an API key, this announcement does not lower your bill.

Context. This follows OpenAI’s 30 July cuts on the cheaper tiers: Luna down 80% ($1/$6 → $0.20/$1.20) and Terra down 20% ($2.50/$15 → $2/$12). Sol held its price through that round. It is not holding now.

The comparison OpenAI wanted you to make

Frontier model	Input	Output	Rate type
Anthropic Claude Fable 5	$10	$50	List
Anthropic Claude Opus 5	$5	$25	List
OpenAI GPT-5.6 Sol	$4	$20	Promotional to 21 Nov

Sol’s input is now 40% of Fable 5’s, its output less than half, and — the point of the exercise — it sits below Claude Opus 5 on both axes. Anthropic launched Opus 5 in July explicitly as a value play at half of Fable 5’s rate. Six weeks later it has been undercut.

Reuters notes OpenAI cited competition from both Anthropic and Chinese models as the driver. Anthropic, per reporting on the response, has not matched with an equivalent discount; its executives have instead been telling corporate customers not to throttle usage over cost. That is a confidence posture, not a price move, and it is a reasonable one — Ramp’s August index showed Anthropic still leading on enterprise account share.

The comparison OpenAI didn’t

Here is the same market sorted by a column vendors do not put on their pricing pages.

Model	Input / Output	Expires
OpenAI GPT-5.6 Sol	$4 / $20	21 Nov 2026
Google Gemini 3.7 Flash	$0.75 / $3.75	31 Dec 2026 → then $1.50/$7.50
DeepSeek V4-Flash	$0.22 / $0.66 off-peak	Every peak window, daily
Anthropic Claude Opus 5	$5 / $25	No stated expiry
OpenAI GPT-5.6 Luna	$0.20 / $1.20	No stated expiry

Three of the five most-quoted rates in this market are conditional. Gemini 3.7 Flash shipped on 13 August at half price with a hard-dated doubling on 1 January. DeepSeek’s cheap rate is only cheap outside its peak hours, which fall inconveniently across European working time. And Sol’s new number is a 92-day window.

This is not an accusation of bad faith. Time-boxing a cut is a rational instrument: it buys share and generates usage data without permanently repricing a tier, and — as AI companies move toward public markets and the scrutiny that brings — without signalling that frontier margins have structurally collapsed. A promotion is reversible in a way a list-price cut is not.

But it changes what a price is for the buyer. “Our frontier cost is $20 per million output tokens” is no longer a fact about your cost base. It is a fact about the next twelve weeks.

We got this wrong eight days ago, and that is the evidence

On 14 August we published a map of this market and concluded: “For frontier reasoning, Claude Opus 5 is the value pick against Fable 5 and GPT-5.6 Sol.”

That was accurate on 14 August and it is wrong today. Sol at $4/$20 beats Opus 5 at $5/$25 on both axes.

We are not flagging this for the correction — it is a routine refresh, and that article’s own advice was to treat the current leader “as a tenant, not a landlord.” We are flagging it because eight days is shorter than any procurement cycle in existence. A mid-size company that started a vendor evaluation the morning that article published would still be in it. The conclusion it started with has already expired, and the conclusion it will finish with is scheduled to expire on 21 November.

When the answer to “which frontier model is cheapest” changes faster than the process designed to answer it, the process is asking the wrong question. Any architecture whose economics depend on a specific vendor being cheapest is now carrying a risk that resolves on a calendar you do not control.

What the discount is actually worth

Concretely, on a team spending $8,000/month on Sol with a coding-agent-typical mix of roughly 30% input and 70% output tokens:

Blended saving ≈ 29% → about $2,320/month
Over the full 92-day window → roughly $7,000

That is real money and it is also, for most teams, less than the fully-loaded cost of a migration — porting prompts, re-tuning agent scaffolding, re-running evaluations, absorbing the regressions you find and the ones you do not. If you are on Opus 5 and considering a switch to capture this, run that arithmetic first. The saving is bounded by a date; the engineering cost is not recovered if the rate reverts.

Where the discount does change decisions:

Workloads you had already benchmarked as Sol-competitive and shelved on cost. The evaluation work is done; the only thing that was missing was the price. Turn them on.
Net-new services with no incumbent to migrate from. No switching cost, so the discount is pure upside — just pick with the November number in your model, not the August one.
Bulk offline jobs you can schedule. Backfills, re-indexing, batch evaluation. If a job can run before 21 November, running it before 21 November is free money.

What it should not change is your default routing for production traffic that is working, unless Sol was already winning your evaluations.

The durable move

Three things, in order of how much they matter.

Put the expiry date in the cost model. Not in a comment — in the model, as a field next to the rate. If your finance spreadsheet says “Sol: $20/M output” without “until 21 Nov,” it is wrong. Budget at the reversion rate and let the discount land as favourable variance. This costs nothing and prevents the specific failure of having spent a windfall you assumed was structural.

Keep a second frontier model warm. Not as a written-down contingency but as a routing target that passes your evaluation suite today. The trigger to switch is now as likely to be a price change as a capability change, and both arrive with days of notice. This is the same argument we made about the routing layer itself getting bought — the mechanics of switching are cheap right now, and the time to build the muscle is while they still are.

Measure blended cost per completed task, not cost per million tokens. A per-token price cut is only a saving if your token consumption stays flat, and on agentic workloads it does not — models differ enormously in how many reasoning tokens they burn to finish a job and how often they need a retry. A model 20% cheaper per token that takes 40% more tokens to complete the same task is a price rise. This is also the only metric that survives a vendor changing its pricing structure rather than its price, which — between long-context surcharges, cached-input tiers, surge windows and latency-tiered endpoints — is now happening more often than straight repricing.

Honest caveats

The reversion is an assumption, not a fact. OpenAI has not said what happens on 21 November. It may extend, or make the rate permanent, and its stated position on earlier cuts has been that efficiency gains funded them. We recommend budgeting at $5/$30 because that is the conservative planning position, not because reversion has been announced.
These are standard short-context rates. Sol carries a long-context surcharge above roughly 272K tokens. If your workloads run long, the headline numbers understate your actual per-call cost and the cut’s effect on it.
Anthropic’s non-response is from reporting, not a statement we can link to. The characterisation that Anthropic is holding price and encouraging continued usage comes from secondary coverage; treat it as such.
DeepSeek’s figures are converted from yuan and move with the exchange rate, and its off-peak/peak split changed again on 23 August. Check the live page before budgeting against it.
No benchmark claim is being made here. This piece is about price. Whether Sol or Opus 5 does your job better is an evaluation question, and the whole argument above is that you should already have the answer before a price moves.

The verdict

OpenAI has taken the frontier price crown from Anthropic, and stamped a date on it.

For most teams the correct action this week is administrative rather than architectural: add the expiry date to the cost model, budget at the old rate, and turn on the specific workloads that were already Sol-shaped and blocked only on price. Do not re-platform a working system to capture twelve weeks of a 29% blended saving.

The larger read is that “which model is cheapest” has become a question with a shelf life measured in weeks, and the labs now have a reversible instrument for changing the answer. That does not make price unimportant — it makes portability the thing price is actually buying you. Our DeepSeek V4-Pro vs Claude Opus 4.8 breakdown and the best AI coding tools guide both assume you can move; ChatGPT, Claude and OpenAI Codex are all worth keeping evaluated rather than picking one and closing the file. Build so that the next expiry date is somebody else’s problem.

Frequently asked questions

What exactly changed in GPT-5.6 Sol's pricing?

For standard short-context use, input dropped from $5.00 to $4.00 per million tokens, cached input from $0.50 to $0.40, and output from $30.00 to $20.00. That is a 20% cut on input and a 33% cut on output, which is why the headline reads 'more than 20%' — the two numbers are not the same reduction. The new rate applies to the pay-as-you-go API, to Codex credits, and is rolling out across eligible ChatGPT Work plans. Pro, Plus and Business subscription pricing is unchanged, so if you consume Sol through a seat rather than a key, nothing about your bill moves. OpenAI's own announcement dates the promotion 21 August to 21 November 2026.

Is GPT-5.6 Sol now cheaper than Claude Opus 5?

Yes, on published list rates, for the duration of the promotion. Sol at $4/$20 sits below Claude Opus 5 at $5/$25 on both input and output — the first time OpenAI's flagship has undercut Anthropic's on both axes simultaneously. Claude Fable 5 remains the most expensive frontier model at $10/$50, so Sol's input is now 40% of Fable 5's and its output is less than half. Two caveats keep this from being a straight swap. First, list price is not spend: output-token efficiency, reasoning-token overhead and retry rates differ by model and can invert a per-token advantage on a real workload. Second, the comparison has a date on it. Opus 5's $5/$25 is a standing rate; Sol's $4/$20 is scheduled to end. Comparing a promotion to a list price and declaring a winner is how you end up re-migrating in November.

What happens on 21 November 2026?

Nobody outside OpenAI knows, and the announcement is deliberately silent on it. There are three plausible outcomes: the rate reverts to $5/$30, the discount is extended, or the lower rate is made permanent because serving costs came down enough to support it. OpenAI has framed previous cuts as efficiency-funded rather than loss-funded, which argues for permanence, and the July reductions on Terra and Luna were not presented as time-boxed. Against that, the company chose to time-box this one specifically, and companies that expect to hold a price do not usually put an end date on it. The planning-safe assumption is reversion: budget at $5/$30, spend at $4/$20, and treat the difference as a windfall rather than as headroom you have already allocated.

Should we migrate our coding agents to Sol to capture the discount?

Only if Sol already wins on your evaluations at the old price. A three-month discount is not enough to justify a migration whose engineering cost you will pay once and whose savings expire. Run the arithmetic before the migration, not after: take your actual monthly frontier-model spend, apply the 20/33% reduction across your real input-output mix, and multiply by three months. If that total does not comfortably exceed the engineering time to port prompts, re-tune agent scaffolding, re-run your evaluation suite and absorb the regressions, the discount is not paying for the work. Where the discount genuinely does change the decision is at the margin — workloads you had already benchmarked as Sol-competitive but shelved on cost, and net-new services with no migration cost at all.

Why are so many AI prices temporary now?

Because the labs are competing on a number they can move in an afternoon, at a moment when none of them can hold a capability lead for long. A promotional rate buys share and generates usage data without permanently repricing the tier or signalling to investors that frontier margins have structurally collapsed — a distinction that matters more as IPO scrutiny arrives. It is also reversible, which a list-price cut effectively is not. Google did the same thing with Gemini 3.7 Flash at $0.75/$3.75 through 31 December, doubling to $1.50/$7.50 on 1 January. DeepSeek did a variant with time-of-day surge pricing. The result is that the phrase 'our per-token cost' has quietly become a statement with a footnote, and buyers who track only the headline number are budgeting on a figure the vendor has already scheduled to change.

How should we actually contract or budget around this?

Four things, none of which require renegotiating anything. Record the expiry date next to every rate in your cost model, so a price and its shelf life travel together — three of the five prices most teams quote today have one. Budget at the post-promotion rate and let the discount show up as favourable variance rather than as capacity you have already spent. Keep at least one alternative frontier model green in your evaluation suite at all times, so switching is a routing decision rather than a project; the same logic applies whether the trigger is a price rise or a capability regression. And measure blended cost per completed task rather than cost per million tokens, because that is the number a per-token cut is supposed to move and the only one that survives a vendor changing its pricing structure rather than its price.

Sources

Related tool reviews

Questions or corrections? Email Pick Right. Want the full list? See all news.