Price war is the only honest description of what happened on the afternoon of 22 September 2026. Anthropic released Claude Opus 5.5, claiming roughly 40 per cent lower running costs than Opus 5. Hours later, OpenAI added two tiers to its GPT-6 family — Sol and Luna — with API prices cut by 50 per cent or more against their GPT-5.6 equivalents. Neither release was a surprise. Both landing the same day was not a coincidence, and a price war fought in public rarely is.
What makes this a genuine price war rather than routine discounting is that an OpenAI spokesperson confirmed to VentureBeat that the new Sol and Luna rates are permanent list prices, not promotional introductory pricing. A promotion expires. A list price has to be defended, and defending a price war position means the cost structure underneath actually moved.
There is a second story sitting awkwardly on top of the first. These are the first frontier releases since Anthropic’s own chief executive called publicly for an industry-wide slowdown on advanced AI development, a call Sam Altman and Elon Musk both joined. This article covers the numbers, what each model is actually for, the open-weight pressure driving the cuts, and how to work out what any of it does to your own bill.
Table of contents
- What the Price War Shipped on One Afternoon
- The Price War in Numbers
- What Claude Opus 5.5 Brings to the Price War
- What GPT-6 Sol and Luna Bring to the Price War
- Why the Price War Started Now
- The Slowdown Call This Price War Answers
- Token Efficiency Is the Real Price War Weapon
- What the Price War Means for Your Bill
- Who Wins and Who Loses in an AI Price War
- Reading the Benchmark Claims During a Price War
- Frequently Asked Questions About the AI Price War
- References
What the Price War Shipped on One Afternoon
Three models, two labs, one day.
Claude Opus 5.5
Anthropic’s new flagship in the Claude 5.5 family. It is priced at $4 per million input tokens and $20 per million output tokens, against $5 and $25 for Opus 5. Cache reads fall from $0.50 to $0.20 and cache writes from $6.25 to $5. It is available through Claude, AWS, Google Cloud and Microsoft Azure.
GPT-6 Sol
A tier below OpenAI’s Astra flagship, aimed at complex repeated work: building features, reviewing code, debugging and analysing data. Sol lists at $2 per million input tokens and $10 per million output, down from $4 and $20 for GPT-5.6 Sol.
GPT-6 Luna
Aimed at high-volume tasks such as extracting information or summarising documents. Luna lists at $0.10 per million input tokens and $0.50 per million output, down from $0.20 and $1.20 — an output cut of roughly 58 per cent, slightly deeper than the headline 50 per cent.
Where you can use them
The GPT-6 models are available in ChatGPT Work and Codex across Plus, Pro, Business and Enterprise tiers, with free users getting Luna only. Opus 5.5 ships with a zero data retention option and EU AI Act watermarking.
The Price War in Numbers
List prices are easy to quote and hard to reason about. The useful exercise is to price the same job on each.
| Model | Input per 1m | Output per 1m | Replaces |
|---|---|---|---|
| Claude Opus 5.5 | $4.00 | $20.00 | Opus 5 at $5 / $25 |
| GPT-6 Sol | $2.00 | $10.00 | GPT-5.6 Sol at $4 / $20 |
| GPT-6 Luna | $0.10 | $0.50 | GPT-5.6 Luna at $0.20 / $1.20 |
| Opus 5.5 cache read | $0.20 | n/a | Opus 5 at $0.50 |
| Opus 5.5 fast mode | $8.00 | $40.00 | New, 2.5x speed |
Pricing a standard job
Take a task that consumes one million input tokens and produces 100,000 output tokens — a realistic shape for document analysis or a long agentic run. Opus 5.5 costs $4.00 plus $2.00, or $6.00. GPT-6 Sol costs $2.00 plus $1.00, or $3.00. GPT-6 Luna costs $0.10 plus $0.05, or $0.15.
The spread is 40 to 1
On that job, Luna is one fortieth the cost of Opus 5.5. That is not a competitive gap, it is a different category of product, and it is why reading a price war off a price table alone tells you almost nothing useful.
Why the cheapest model is not the answer
The cheapest run is the one that gets the job done first time. A model that costs a fortieth as much but needs three attempts and a human correction is not cheaper. That is the whole argument for a tiered line-up, and it is why both labs answered the price war with tiers rather than a single cut.
What Claude Opus 5.5 Brings to the Price War
Anthropic published an unusually detailed benchmark table, and it does not claim a clean sweep.
Where it leads
Opus 5.5 posts 66.4 per cent on Terminal-Bench 4.0 against 57.9 for GPT-6 Astra, 55.8 for Fable 5.1 and 52.3 for Opus 5. It leads on FrontierCode v1.1 at 54.4 per cent, on Humanity’s Last Exam at 67.7 per cent, and on GDPval-AA v2.1 with a score of 1846 against Astra’s 1542.
Where it does not
GPT-6 Astra beats it on AutomationBench, 41.4 per cent to 40.0, and on Terminal-Bench-Science, 64.6 per cent to 58.7. Publishing the two benchmarks you lose is a reasonable signal that the rest of the table is being reported straight.
| Benchmark | Opus 5.5 | GPT-6 Astra | Opus 5 |
|---|---|---|---|
| Terminal-Bench 4.0 | 66.4% | 57.9% | 52.3% |
| FrontierCode v1.1 | 54.4% | 53.3% | 48.0% |
| Humanity’s Last Exam | 67.7% | 57.2% | 63.6% |
| AutomationBench | 40.0% | 41.4% | 26.9% |
| Terminal-Bench-Science | 58.7% | 64.6% | 29.0% |
| GDPval-AA v2.1 | 1846 | 1542 | 1708 |
The safety numbers in the same release
Anthropic reports that Opus 5.5 attempted to circumvent boundaries about 85 per cent less often than Opus 5 across a behavioural audit of nearly 2,000 scenarios, and that it ties Fable 5.1 for the lowest attack success rate on Gray Swan’s prompt injection evaluation. For anyone whose cybersecurity posture has to account for agents reading untrusted text, that second figure is the more consequential one.
The customer-reported results
Deloitte reports Opus 5.5 catching 72 per cent of bugs in code review at a low effort setting, against 56 per cent for Opus 5 at high effort. Hebbia reports 86.6 per cent coverage on end-to-end finance workflows against 60.3 per cent. Box reports answers 40 per cent less verbose without losing accuracy.
What GPT-6 Sol and Luna Bring to the Price War
OpenAI’s two additions are more clearly segmented than Anthropic’s single release.
Sol is the working model
OpenAI positions Sol for complex work that developers and knowledge workers repeat: building features, reviewing code, debugging, analysing data. It reports roughly half the factual errors of its predecessor and says it matches Fable 5.1’s coding performance at a lower cost.
Luna is the volume model
Luna is for high-volume, low-judgement work — extracting fields, summarising documents, classifying records. At $0.10 per million input tokens it is priced to be used indiscriminately, which is exactly the point.
Astra stays on top
GPT-6 Astra remains OpenAI’s flagship for the most demanding workloads. Sol and Luna are cost-effective alternatives beneath it rather than replacements for it, which mirrors the shape of Anthropic’s line-up.
Free users get Luna
Making Luna the free tier is a distribution decision as much as a pricing one. It puts a competent extraction-and-summarisation model in front of everyone at a cost OpenAI can absorb.
Why the Price War Started Now
Two pressures converged, and only one of them gets talked about.
Open weights set the price war floor
CNBC reports both labs facing stiff competition from cheaper open-weight models, naming the Chinese firms Alibaba, Moonshot AI and DeepSeek. An open-weight model you can host yourself sets a hard ceiling on what a proprietary API can charge for commodity work, and extraction and summarisation are commodity work.
Customers started counting
The releases come as both labs work to satisfy customers looking to rein in AI spending. After two years of enthusiastic budget allocation, buyers now have usage data, and usage data produces procurement questions. A price war is the most direct answer a vendor has to a renewal conversation.
Efficiency made the cut affordable
Dianne Penn, Anthropic’s head of product management, research and labs, told CNBC the company is innovating on “how to make that thinking, how to make the answering more efficient, so it uses less tokens depending on your effort setting”. Cheaper per token plus fewer tokens is how a 20 per cent list cut becomes a 40 per cent running-cost claim, and it is how a price war is funded without losing margin.
The price war logic is symmetrical
Neither lab can let the other hold a price advantage on agentic coding, because that is where the highest-value developer workloads sit. Once one moves, the other has to, and both had clearly prepared to move on the same day.
The Slowdown Call This Price War Answers
The context nobody at either lab wanted attached to a pricing announcement.
What was said
Former Anthropic researcher Jacob Coxon posted on 8 September 2026 that he had quit, warning that both labs were “gambling with our lives”. Anthropic chief executive Dario Amodei subsequently called for an industry-wide slowdown on developing advanced AI. Sam Altman and Elon Musk joined that call.
What then happened
Two weeks later both labs shipped more capable models at substantially lower prices. Cheaper capability is, by construction, more widely deployed capability. Whatever a slowdown means, a price war does not appear to be it.
The defensible reading
Neither release claims a capability jump on the scale of a new generation. Opus 5.5 is a point release optimised for efficiency; Sol and Luna are tiers beneath an existing flagship. A lab could argue consistently that it slowed frontier capability work while continuing to make existing capability cheaper.
The uncomfortable reading
Price is the main lever on deployment. Halving the cost of agentic coding does more to increase real-world AI activity in the next quarter than most capability improvements would. Efficiency work is not neutral with respect to how much AI gets used, which is what makes this price war a safety question as well as a commercial one.
Token Efficiency Is the Real Price War Weapon
The list price is the smaller half of the story, and the harder half to verify.
The arithmetic of the 40 per cent claim
Opus 5.5’s input price is 80 per cent of Opus 5’s and its output price is 80 per cent of Opus 5’s. A pure list-price comparison therefore gives a 20 per cent reduction, not 40. The rest comes from the model producing fewer tokens for the same task — which is a claim about behaviour, not about a price list.
Verbosity is now a priced feature
Box’s report of answers 40 per cent less verbose without losing accuracy is a cost saving measured in output tokens. So is Optiver’s report of matching Opus 5’s quality in about half the turns and output tokens. Those are bill reductions that never appear on a price page.
Cache pricing moved the most
Opus 5.5’s cache reads fell from $0.50 to $0.20, which is 40 per cent of the old price — a deeper cut than either the input or output line. Any workload with a large stable prefix, which describes most agentic and document workflows, benefits disproportionately.
Why you must measure your own workload
Two organisations running the same model on different task shapes will see completely different savings from this price war. The only way to know what a price war did to your costs is to price your own traffic before and after.
What the Price War Means for Your Bill
Four practical conclusions, none of which require picking a side.
Re-run your model selection after a price war
Anything you routed to a premium model for reliability reasons six months ago deserves a fresh test against a cheaper tier. Sol matching Fable 5.1 on coding at a lower price, if it holds on your work, changes the default.
Split your traffic by judgement required
The 40-to-1 spread between Luna and Opus 5.5 only pays off if you actually route by task. Extraction, classification and summarisation belong on the cheap tier; anything where a wrong answer costs real money does not. This is the same routing discipline that makes intelligent automation projects economic rather than merely impressive.
Do not assume price cuts continue
A price war is a competitive state, not a trend. Sol and Luna are confirmed as permanent list prices, which is genuinely reassuring, but nothing obliges either lab to keep cutting. Build cost models on today’s prices, not on an extrapolated curve.
Watch efficiency, not just rates
The most durable saving in this release is fewer tokens per task, and it is invisible on a pricing page. Tracking tokens per completed job, rather than cost per million tokens, is the metric that survives the next repricing — and it is the number that tells you which of the current AI models and tools is actually cheapest for your work.
Who Wins and Who Loses in an AI Price War
A price war redistributes value, and it is worth being clear about the direction.
Buyers win, at least for now
Anyone paying for tokens is straightforwardly better off. A 50 per cent cut on Sol and Luna and a 40 per cent running-cost reduction on Opus 5.5 land in the same month, which is an unusually good quarter to be renewing a contract.
Small builders win most
A price war compresses the cost of experimentation. Products that were marginal at GPT-5.6 pricing become viable at Luna’s $0.10 per million input tokens, and a category of application that was previously uneconomic now is not.
The middle tier is squeezed
Providers reselling frontier models with a thin margin have just watched their input costs and their achievable price both fall. A price war between two suppliers is rarely comfortable for the intermediaries between them and the customer.
Open-weight hosts face the same pressure
The open-weight providers that started this price war by undercutting proprietary APIs now face a Luna tier priced at a tenth of a dollar per million input tokens. Self-hosting has to beat that on total cost including operations, which is a much harder argument at these prices than it was a year ago.
The labs are betting on volume
Neither Anthropic nor OpenAI cuts prices out of generosity. A price war is a bet that lower unit prices produce more than proportionally more usage, and that the resulting scale is defensible. That bet has worked in cloud infrastructure for two decades; whether it works when the marginal cost is a GPU-hour rather than a disk is not yet settled.
Reading the Benchmark Claims During a Price War
When both labs publish on the same afternoon, the comparison tables are marketing artefacts as well as research.
Each lab chose its own suite
Anthropic’s table leads with Terminal-Bench 4.0 and FrontierCode. OpenAI’s claim is that Sol matches Fable 5.1 on coding at lower cost. Neither ran the other’s preferred evaluation under the other’s conditions, which is normal and still worth noticing.
Scaffolds move scores more than models do
A computer-use or agentic-coding score depends heavily on the harness: how many steps are allowed, what tools are available, how retries are counted. Two credible parties can report very different numbers for the same model without either being dishonest.
Cost-adjusted performance is the honest metric
Anthropic’s most defensible comparison is that Opus 5.5 matches GPT-6 Astra on coding at roughly 20 to 40 per cent of the cost. In a price war, performance per pound is the number that decides purchases, and it is the one both labs are now optimising.
Your workload is the tiebreaker
None of the published figures describe your repository, your documents or your customers. Running the same fifty real tasks through each tier and counting completions and total spend takes an afternoon and settles the question better than any table.
Frequently Asked Questions About the AI Price War
How much cheaper is Claude Opus 5.5?
List prices fell about 20 per cent, from $5 and $25 per million input and output tokens to $4 and $20. Anthropic’s 40 per cent figure combines that with the model using fewer tokens per task. Cache reads fell furthest, from $0.50 to $0.20.
Are the GPT-6 Sol and Luna prices permanent?
Yes. An OpenAI spokesperson confirmed to VentureBeat that the new rates are permanent list prices rather than promotional or introductory pricing.
Which model is best for coding?
Anthropic reports Opus 5.5 leading Terminal-Bench 4.0 at 66.4 per cent against 57.9 for GPT-6 Astra, while OpenAI says Sol matches Fable 5.1 on coding at a lower cost. The two claims are not directly comparable, so test both on your own repository.
Why did both launches happen on the same day?
Neither lab can concede a price advantage on developer workloads, so a move by one forces a move by the other. Both were clearly prepared, and open-weight competition from Alibaba, Moonshot AI and DeepSeek has been pushing on commodity pricing for months.
Does this contradict the call for an AI slowdown?
It sits uncomfortably beside it. Both labs can argue these are efficiency releases rather than capability jumps, but cheaper models mean more deployed AI, which is not obviously what a slowdown was meant to produce.
References
Anthropic and OpenAI announce more powerful (and cheaper) AI models
Anthropic and OpenAI roll out cheaper models in first release since call for slowdown
OpenAI launches GPT-6 Sol and Luna, boasting lower cost and fewer mistakes
OpenAI cuts GPT-6 prices in half with Sol and Luna
OpenAI, Anthropic Launch Dueling Models in API Price War
What slowdown? OpenAI, Anthropic release dueling models as AI price wars heat up
OpenAI releases GPT-6 Sol and Luna and cuts token prices in half
More AI coverage: explore Progressive Robot's AI Models, Tools & Releases hub — hands-on reviews, setup guides and benchmarks in one place.
