Guide 30 · Comparison
The AI Model Price War of September 2026
OpenAI and Anthropic launched flagship models on the same day and slashed token prices — here is what it means for your API bill.
By Abhit Tomar · Updated November 17, 2026 · 14 min read

Short answer
On September 22, 2026, OpenAI and Anthropic launched new models the same day and cut prices. GPT-6 Sol fell 50% to $2/$10 per million tokens, GPT-6 Luna targets background work at $0.10/$0.50, and Claude Opus 5.5 dropped 20% to $4/$20. Benchmarks differ by lab, so compare prices, not press-release scores.
What happened on September 22, 2026
On September 22, 2026, OpenAI and Anthropic did something the industry rarely sees: both launched new models on the very same day, and both used the occasion to slash API prices. OpenAI shipped two models — GPT-6 Sol, its flagship, and GPT-6 Luna, a cheaper model built for background work. Anthropic answered with Claude Opus 5.5, the latest heavyweight in its Opus line.
Six days later, on September 28, Anthropic kept the pressure on by releasing Claude Sonnet 5.5 at $2 input and $10 output per million tokens, saying it needs fewer tokens to complete the same work — effectively another price cut, reported by Reuters. For a sense of how unusual the timing was, see this breakdown of the same-day launches.
The new prices, in plain English
Token prices are quoted per million tokens. GPT-6 Sol costs $2 per million input tokens and $10 per million output tokens — exactly half of its predecessor's $4/$20. GPT-6 Luna sits far below at $0.10 input and $0.50 output, making it the budget pick for summarization, classification, and other background jobs.
On the Anthropic side, Claude Opus 5.5 is priced at $4 input and $20 output per million tokens — 20% below Opus 5's $5/$25. Anthropic says that works out to roughly 40% cheaper on typical workloads with 30%+ faster output. Claude Sonnet 5.5, the mid-tier release, lands at $2/$10 — matching GPT-6 Sol's headline rates. For reference, xAI's Grok 4.7 charges $2 input ($0.50 cached) and $6 output for prompts under 200k tokens.
Output tokens usually cost about five times input tokens, so generation-heavy apps feel these cuts the most. To see what the new numbers do to a real bill, run them through our AI token price comparison tool or the token calculator before your next launch.
A useful rule of thumb for reading these numbers: a million tokens is roughly 750,000 words of English text, so a million input tokens is a very large pile of reading and a million output tokens is a very large pile of writing. That is why the input-output split matters so much — an app that reads a lot and writes a little pays mostly the cheap input rate, while a chatbot that generates long answers pays mostly the expensive output rate. Cached input is a third category worth knowing: when you send the same prompt prefix repeatedly, some providers charge a fraction of the normal input price for the repeated part, which is why Grok 4.7 lists a separate $0.50 cached rate. Before comparing headline prices, work out your own input-output mix; two models with identical sticker prices can cost very different amounts on your workload.
What the cuts mean: builders vs casual users
- If you build on APIs: a 50% input/output cut on GPT-6 Sol, or the 20% Opus 5.5 cut, can change a product's unit economics overnight. Re-run your pricing math with the AI token price comparison tool — the cheapest model on your shortlist may have changed.
- If you run background workloads: GPT-6 Luna at $0.10/$0.50 is explicitly priced for jobs that used to be too expensive to automate — document summarization, bulk classification, draft generation.
- If you are a casual user: API prices do not touch you directly, since you use subscriptions. But cheaper tokens let app builders offer more generous free tiers, so expect better free AI products.
- If you are choosing a model: price is only half the story. Check our full model comparison and the Muse vs ChatGPT vs Claude guide before deciding.
An honest caveat: you cannot compare these benchmarks
Every launch came with impressive benchmark scores. OpenAI reported that GPT-6 Sol scored 33.2% on AutomationBench at roughly $0.27 per task. Anthropic countered that Claude Opus 5.5 scored 89.9% on SWE-bench Pro and 66.4% on Terminal-Bench 4.0.
Here is the honest problem: no two labs' benchmarks measure the same thing. AutomationBench, SWE-bench Pro, and Terminal-Bench 4.0 test different skills in different ways, so any cross-vendor score comparison is apples-to-oranges. Treat every vendor chart as marketing with a p-value — these numbers are vendor-reported via press, not independently verified.
Read benchmarks as directional, not decisive. They tell you which skills each lab wanted to brag about, not which model is best for your task. The only benchmark that matters is your own: run your prompts through each candidate model, check the quality yourself, and pick the cheapest one that passes. Our full model comparison is a good starting point.
The honest alternative to vendor benchmarks is a small test of your own. Take ten or twenty prompts from your real workload — the actual questions, documents, or tasks your users send — and run the same set through each candidate model. Score the outputs yourself against what good looks like for your product, and divide the total cost by the number of tasks that passed. That cost per completed task is the number that matters, because a cheaper model that fails half your prompts and needs retries can cost more than a pricier one that gets it right first time. You do not need a lab or a framework for this; a spreadsheet, an afternoon, and your real prompts are enough.
What to watch next
Price wars rarely stop after one round. The open question is whether xAI, Google, or Meta respond with cuts of their own — and whether the September 28 Sonnet 5.5 launch pushes OpenAI to trim GPT-6 Sol or Luna further. We will keep tracking the moves here at this independent guide hub.
Watch your actual invoices, not press releases. If you run production workloads, set a date in October to re-price your stack with the token calculator — the savings only land when you redeploy on the cheaper model, not when the announcement goes out.
One specific thing to watch is whether the September 28 Sonnet 5.5 claim holds up in practice: Anthropic said it needs fewer tokens to complete the same work, which would make its effective price lower than the $2/$10 sticker suggests — but only if your prompts behave the way theirs did. The other is whether the price war spreads beyond the two labs that started it. Google and Meta both run frontier models with their own pricing, and xAI's Grok 4.7 was already sitting at $2/$6 when the war began. If any of them answers with a cut of their own, the floor moves again; if none does, September's prices may simply become the new normal.
What actually drove prices down
The September price cuts did not come out of nowhere; several forces converged at once. Compute costs per unit of model capability have been falling as hardware improves and inference infrastructure gets more efficient, so providers can serve the same quality for less. Competition intensified as multiple labs shipped capable models within weeks of each other, and price became the most visible way to win developers and enterprise contracts. Open-weight models added pressure from below: when a good model is available nearly free, commercial APIs have to justify their price with reliability, features, or support rather than raw capability. Finally, several providers appeared to be pricing aggressively to gain market share, accepting thinner margins now in hopes of locking in users later. Any one of these would have nudged prices down; together they produced a genuine price war.
How it affects everyday users
For casual users, the price war mostly shows up as better free tiers and cheaper subscriptions. Providers competing for attention tend to raise usage limits, add features to lower tiers, and make premium capabilities available at prices that would have seemed unrealistic a year ago. If you use AI tools occasionally — drafting help, homework questions, quick research — you are likely getting more for your money than ever, and switching between providers costs you almost nothing. The less obvious effect is on the tools built on top of these models: cheaper API costs let app developers add AI features without charging much, so expect more AI sprinkled into everyday software. The main thing to watch is quality consistency; when providers cut prices, some quietly adjust rate limits or route traffic to cheaper infrastructure, so if a tool you rely on gets flaky, price pressure upstream is worth suspecting.
Comparing value across tiers and whether cheap prices last
Comparing plans during a price war means looking past the headline number. Check what you actually get: message or token limits, which model you are served (flagship or a smaller variant), priority during peak hours, and whether features like file uploads, web browsing, or memory are included. A cheaper plan on a weaker model can be worse value than a pricier plan on the flagship if your work needs the stronger one — but for routine tasks the smaller models are often indistinguishable, and that is where the savings are real. As for whether prices stay low: some of this is structural, since efficiency gains are real and open models keep a ceiling on what anyone can charge. But aggressive below-cost pricing to win share is not sustainable forever, and history suggests prices stabilize once the market consolidates. Expect the floor to stay lower than it was, with occasional promotional dips, rather than a permanent race to zero.
How to pick a plan during the volatility
Volatility rewards flexibility, so avoid long commitments while prices are moving. Month-to-month billing costs a little more per month but lets you switch when a better deal appears; annual plans lock in today's price, which is only smart if you are confident it is near the bottom. Start by measuring your actual usage for a month before choosing — most people overestimate how much they use, and a cheaper tier often covers it. If you use multiple providers, keep one primary subscription and use free tiers for the rest rather than stacking paid plans. Revisit your choice quarterly: set a reminder, check whether your usage or the market has changed, and switch without sentimentality. The providers are competing for you; let them.
How to estimate your own savings
Start with your provider dashboard, which shows last month's token usage split into input and output. Multiply the input total by the old input rate and the new one, do the same for output, and the difference is your monthly saving — if you switch models and nothing else changes. Two things commonly spoil the estimate. First, cached input and batch pricing can already be discounting your current bill, so compare against what you actually paid, not the sticker rate. Second, switching models changes how many tokens a task needs: a model that writes more verbosely costs more output tokens per answer. Run the new model on a sample of real traffic first, measure the tokens it actually uses, and only then project the saving.
| Model | Input / 1M tokens | Output / 1M tokens |
|---|---|---|
| GPT-6 Sol | $2 | $10 |
| GPT-6 Luna | $0.10 | $0.50 |
| Claude Opus 5.5 | $4 | $20 |
| Claude Sonnet 5.5 | $2 | $10 |
| Grok 4.7 | $2 ($0.50 cached) | $6 |
Keep reading
02 · ComparisonChatGPT Dots vs Grok Bot vs Muse AI: Which Always-On Agent Fits You?Three always-on agents launched within seven weeks of each other. What each one actually does, what it costs, and which one matches your week.Read guide
06 · ComparisonMuse AI vs ChatGPT vs Claude vs Meta AI: Which Assistant Fits Your Work?Four assistants, four different jobs: who finishes tasks, who writes best, who costs what — and which one you should actually use.Read guide
07 · ComparisonMuse AI vs Claude: Which Fits Your Workflow?Choose based on the work you repeat, not a generic ranking.Read guide