Both AI Labs Cut Prices on the Same Day — And Your Plan Gets More Room

TL;DR: Anthropic and OpenAI both cut prices on 22 September, hours apart. Claude Opus 5.5 costs 40% less to run than Opus 5 and matches Fable 5.1 on most work — and Pro, Max and Team plans get bigger five-hour limits, which is the part that actually reaches you. OpenAI halved token prices with GPT-6 Sol and Luna. After a month of usage-limit anxiety, this is the first genuinely good pricing news — with the usual caveat that nobody has compared the two new models head to head, because they launched the same afternoon.

What Anthropic shipped

Claude Opus 5.5 arrived on 22 September. Anthropic’s own summary: it “performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5.”

The per-token prices for Claude Opus 5.5, from Anthropic:

Per million tokens	Opus 5.5	Opus 5 (previous)
Input	$4	$5
Output	$20	$25
Cache reads	$0.20	$0.50
Cache writes	$5	$6.25

The headline per-token cut is 20%. The “40% drop” is for typical workloads, where the much cheaper cache reads do most of the work. There’s also a Fast mode at $8 input / $40 output if you want speed over cost.

The line that matters most for subscribers isn’t in the pricing table. Anthropic says Pro, Max, Team and seat-based Enterprise plans get increased five-hour usage limits. You don’t pay per token on those plans — so a cheaper model reaching you looks like more room, not a smaller bill.

Anthropic also cites one tester completing “a 680,000-line code migration in less than a day”, and a 1846 Elo on its knowledge-work evaluation. Worth pairing with Anthropic’s own unusually frank caveat: “at these levels of capability we’ve found that benchmark margins have become a less reliable guide to real-world differences.”

What OpenAI shipped hours later

The same day, OpenAI released GPT-6 Sol and Luna, and per VentureBeat cut prices by half or more:

Per million tokens	New price	Previous
GPT-6 Sol	$2 / $10	$4 / $20
GPT-6 Luna	$0.10 / $0.50	$0.20 / $1.20

OpenAI confirmed these are “permanent prices, not promotional or introductory pricing” — which matters, because introductory pricing is how this usually goes.

Sol lands at exactly half the cost of Claude Opus 5.5 per token. OpenAI claims Sol beats Claude Opus 5 on a workflow benchmark at a fraction of the cost per task — note that’s Opus 5, the previous model, not Opus 5.5. VentureBeat is direct about why: “direct Sol-versus-Opus 5.5 performance data [is] unavailable.” They launched hours apart. Nobody has run them against each other yet.

VentureBeat adds its own caution, which we’d endorse: “vendor comparisons warrant caution. Different labs can produce different relative results when benchmarks are run under different configurations.”

What Claude Opus 5.5 changes about the advice we gave on Monday

Three days ago we wrote about Claude Code Projects, which fans your work into parallel cloud sessions. The warning was specific: every thread defaults to Opus at high effort, which Anthropic’s own documentation called the setting that “draws on your plan fastest.”

That advice needs updating, and in a good way. The expensive default just got 40% cheaper to run, and the plans it runs on just got bigger five-hour limits. The same feature, used the same way, should now cost you meaningfully less of your allowance.

What hasn’t changed: there’s still no cap on how many threads run at once beyond 200 new ones a day, a thread that hits your limit still waits and resumes into your next window, and idle threads watching a pull request still wake up and spend. Cheaper per unit is not the same as bounded. Check Project settings → General before your next run regardless — you’re now choosing between cheaper options, not avoiding a trap.

Why these price cuts matter if you don’t code

Two honest framings, because it’s easy to overstate this.

The thing that reaches you is limits, not prices. If you pay £17 or £200 a month, you never see a per-token price. What you see is how much work fits before you hit a wall. A model that costs 40% less to run means your subscription buys more of it — and Anthropic has said the five-hour limits go up. That’s the mechanism, and it’s real.

But this is a price war, and price wars have a middle. Two labs cut on the same day is competition working, and it’s the first pricing news this month that isn’t a lawsuit, a pause or a warning. We covered a class action over what Claude Max’s “20x” claim describes and OpenAI pulling its $200 tier off sale within days of each other. Against that, a genuine cost reduction is welcome.

The caution is that cheaper tokens don’t fix the thing that made those stories possible: you still can’t see what a unit of your work costs, so you still can’t tell whether a 40% cut means you get 40% more done. The only way to know remains measuring your own usage before and after.

One thing worth knowing about Claude Opus 5.5

Anthropic documents something that surprised a few people: safeguards can route your request to an older model. In its words, “Most cybersecurity tasks will be re-routed to Opus 4.8”, and biology requests over a threshold go elsewhere.

This has been reported as requests being “secretly” downgraded. That framing isn’t quite fair — Anthropic published it on the launch page — but the practical point stands: if you’re working on something security-adjacent, you may be talking to a different, older model than you think, and mid-conversation. For most people building an app, this will never come up. If you’re doing anything touching security tooling, it’s worth knowing why the answers might feel inconsistent.

Who should care (and who shouldn’t)

• Using Claude Code heavily, especially Projects: the most direct beneficiary. More headroom on the same plan, and the expensive default is now less expensive.
• On Pro and hitting limits: the increased five-hour limits are aimed at you. Measure a normal session again before assuming you need to upgrade.
• Paying per token through an API: the real winner, on either side. Sol at $2/$10 and Luna at $0.10/$0.50 are steep cuts, and OpenAI says they’re permanent.
• Building with Lovable, Bolt or Replit: nothing changes today. Those tools set their own prices and absorb model costs; whether they pass savings on is their decision, and worth watching over the next month.
• Choosing between Claude and ChatGPT on price alone: don’t, yet. No one has compared the new models head to head, and both labs caveat their own benchmarks.

Our take

We’ve spent this month writing sceptical pieces about AI pricing, so it’s only fair to say plainly: this is good news, and competition is why.

Two labs cutting on the same day, with OpenAI explicitly saying its prices are permanent rather than introductory, is the market working the way buyers want it to. And Anthropic passing savings to subscribers as bigger limits rather than only to API customers is the right call — most of our readers are on a plan, not a meter.

What we’d resist is treating the benchmark claims as the story. OpenAI’s headline comparison is against the previous Claude model, because the current one launched the same afternoon. Anthropic says outright that benchmark margins are now a poor guide to real-world differences. Both companies have, in effect, told you not to pick on the numbers — and they’re right.

So take the cheaper prices, take the bigger limits, and judge the models the only way that has ever worked: run your own real task on both and see which one finishes it. That answer arrives in an afternoon and doesn’t need a leaderboard.

Not sure which tool — or which plan — fits what you’re building? Take the 60-second Vibe Coding Tool Finder quiz →

FAQ