Yesterday was the most entertaining day in AI pricing that I’ve seen in a long time. In the span of about 90 minutes, Anthropic released Claude Opus 5.5 and OpenAI released GPT-6 Sol and GPT-6 Luna. Four flagship-grade models, two rival labs, one shared message: faster, cheaper, better.
I’ve watched AI pricing for years and I honestly can’t remember a day like this. Two of the biggest labs on the planet effectively launched a price war against each other, on the same afternoon. And buried under all the marketing, there’s something genuinely important here for anyone building an AI startup: the economics of running AI features just shifted, possibly by a lot.
So let’s do what I’d want someone to do for me. What actually launched. What it costs, in real dollars. What the catch is (there’s always a catch). And the practical playbook for what a founder should do about it this week.
What Actually Launched on September 22
Let’s separate the news from the noise, because the announcements are fresh and the facts matter more than the hype.
Anthropic: Claude Opus 5.5
Claude Opus 5.5 is the first model in Anthropic’s new Claude 5.5 family, and it’s positioned as the new top of the Claude lineup. Anthropic says it performs at the level of Claude Fable 5.1 — the creative/narrative flagship from earlier this month — on most work, while costing roughly 40% less to run than its predecessor, Opus 5, on typical workloads.
The headline numbers: $4 per million input tokens and $20 per million output tokens, down from $5 and $25 on Opus 5. But here’s the detail that actually matters most for agentic workloads — cache reads dropped to $0.20 per million tokens, 60% less than Opus 5’s $0.50. Cache reads are where the real money goes in coding agents and long-running workflows, because every round trip re-reads the context. That’s the line item to watch.
Anthropic also claims output generation is more than 30% faster, and there’s a “fast mode” in Claude Code and the Claude Platform (up to 2.5x speed) priced at $8/$40 per million input/output tokens. Early tester stories include a 680,000-line code migration completed in under a day, and one case where Opus 5.5 audited a 200,000-line codebase in under 3 hours where Opus 5 took over 20 hours and 2.5x the tokens. Those are vendor-reported numbers, so take them with the usual grain of salt — but the direction is hard to miss.
One delicious bit of context: this is Anthropic’s first model release since CEO Dario Amodei called for “pacing the frontier” — the idea that AI progress should slow down so safety practices can keep up. Then Anthropic shipped a faster, cheaper flagship the following week. The irony was not lost on anyone, and frankly it’s a perfect snapshot of how this industry works right now. Reuters covered the launch here.
OpenAI: GPT-6 Sol and GPT-6 Luna
About 90 minutes after Anthropic’s announcement, OpenAI expanded the GPT-6 family (which debuted with GPT-6 Astra earlier this month) with two new tiers:
- GPT-6 Sol — $2 per million input, $10 per million output. Positioned for coding, complex reasoning, and professional work. That’s half the promotional price of GPT-5.6 Sol. On OpenAI’s own factuality evaluation, Sol makes about half as many errors as its predecessor and approaches Astra-level quality on some tasks.
- GPT-6 Luna — $0.10 per million input, $0.50 per million output. The everyday workhorse for high-volume, routine tasks: summarization, extraction, classification, quick Q&A. Per OpenAI, high-effort Luna roughly matches GPT-5.6 Sol quality at around one hundredth of the cost.
For reference, GPT-6 Astra — the flagship — sits at $10/$50 per million tokens. So OpenAI now has a clear three-tier lineup: Astra for the hardest work, Sol as the coding workhorse, Luna as the volume play. Both new models are API-only (no self-hosting) and are already live in the API as gpt-6-sol and gpt-6-luna, plus ChatGPT Work and Codex. Reuters has the details on the pricing, and MarkTechPost ran through the benchmarks.
Oh, and one housekeeping note OpenAI slipped in: GPT-5.5-instant retires on October 14, 2026, which means default models in ChatGPT will shift. If your product wraps ChatGPT defaults, double check what changes for you.
The Price Sheet, All in One Place
Prices are scattered across four announcements, so here’s the whole board as of this week (per million tokens):
- Claude Opus 5.5: $4 in / $20 out / $0.20 cache reads (fast mode: $8/$40)
- GPT-6 Astra: $10 in / $50 out (flagship tier)
- GPT-6 Sol: $2 in / $10 out
- GPT-6 Luna: $0.10 in / $0.50 out
- Claude Opus 5 (previous gen): $5 in / $25 out / $0.50 cache reads
- GPT-5.6 Sol (previous gen promo): $4 in / $20 out
Real talk: input-token prices are the least important number here. For most AI startups, the money is in output tokens (the model writing code, writing copy, reasoning through problems) and in cache reads (agents re-reading context every step). Opus 5.5’s cache-read cut and Luna’s $0.10/$0.50 pricing attack exactly those cost centers. That’s why this launch wave matters more than the usual “we made it 10% better” release.
Why This Is Happening Right Now
It’s worth asking why both labs suddenly decided that cheaper was the story. A few things are colliding at once.
First, inference got cheaper to run. Better caching, more efficient serving, and architectural improvements mean the labs’ costs dropped — and they’re passing some of that through. OpenAI said it outright: “Improvements in caching and inference let us serve these models at lower cost, and we’re passing those savings directly on to users and customers.”
Second, open-weight models are breathing down everyone’s neck. Earlier this month DeepSeek open-sourced a native vision model under an MIT license, Qwen’s snapshots are topping coding leaderboards, and there’s a real, growing chunk of developers choosing open weights purely on cost. When the free-ish alternative gets good enough, the paid options have to move.
Third — and this is the spicy one — it’s just competition. Nobody wants to be the expensive lab. When Anthropic cuts 40% and OpenAI responds 90 minutes later with half-price workhorse models, that’s not coincidence. That’s two companies watching each other’s pricing page. For founders, this is the rare case where corporate rivalry directly fills your pockets.
“Improvements in caching and inference let us serve these models at lower cost, and we’re passing those savings directly on to users and customers.” — OpenAI, on the GPT-6 Sol and Luna launch
That one sentence is basically the thesis of this whole article. The price war is real, and the beneficiary is whoever ships products on top of these models. That’s you.
What This Actually Means for Your Startup
Okay, enough industry commentary. Let’s talk about what changes for someone actually building.
1. Your AI margins just got a re-rate
If you run AI features on any of the previous-gen models, go look at your actual spend this week. A 20–50% cut in per-token pricing on your heaviest workloads is not a rounding error — for a lot of AI SaaS products, inference is the single biggest COGS line. Our breakdown of what it costs to start an AI business treats API spend as a major budget line for exactly this reason. Yesterday’s launches could move you from “this feature barely breaks even” to “this feature is profitable.”
2. The era of one default model is over
Here’s the thing the labs are really selling: a portfolio, not a product. OpenAI’s three-tier GPT-6 family (Astra / Sol / Luna) and Anthropic’s Opus/Sonnet/Haiku structure all push toward the same architecture — route each task to the cheapest model that handles it well. Luna for classification and extraction. Sol for coding. Opus 5.5 or Astra for the genuinely hard stuff. If your product calls one model for everything, you’re leaving real money on the table.
3. Features that were too expensive are now viable
This is my favorite part. Every product team has a list of AI features that were killed by unit economics — per-user summarization, background agents, full-document analysis, anything that burns thousands of tokens per action. When Luna-class models cost $0.10 per million input tokens, a huge chunk of that “too expensive” list becomes “shippable this sprint.” I’d honestly go re-open that list.
4. Your competitors got the same memo
The flip side. Your AI bill went down, and so did everyone’s. If your moat is “we use the good models,” you don’t have a moat. The differentiator is getting further away from the model layer: your data, your workflow, your distribution. This pairs with the broader strategy in our guide on starting an AI startup — the model is infrastructure now, and infrastructure is priced like a commodity.
The Founder’s Playbook: Actually Cutting Your AI Bill
Prices dropping doesn’t automatically save you money — most founders I see waste more through sloppy architecture than they save through cheaper models. Here’s the playbook I’d run this week.
Step 1: Route tasks by difficulty, starting today
Audit where your tokens actually go. In almost every AI product, 70–80% of calls are easy work: classification, extraction, summarization, formatting, routing. Those belong on Luna-class models ($0.10/$0.50) or the cheapest capable Claude tier — not on a $10/$50 flagship. Reserve Opus 5.5 or Astra for the 20% of tasks that genuinely need frontier reasoning: hard coding, complex analysis, agentic multi-step work.
The rule of thumb: if a human couldn’t tell which model answered, you’re overpaying. Start everything on the cheap tier and promote only the tasks that fail quality checks.
Step 2: Cache aggressively
Opus 5.5’s cache reads at $0.20 per million tokens are 60% cheaper than before — but only if you actually use prompt caching. Any system prompt, codebase context, or document that stays stable across requests should be cached. In agentic loops, this is often the majority of your bill. If you haven’t set up prompt caching on your heaviest workflows, do that before you even think about switching models. It’s free money.
Step 3: Re-run your evaluations on the new models
This is the step everyone skips, and it’s the most important one. The benchmarks in the launch announcements are vendor claims — Anthropic’s own announcement admits “benchmark margins have become a less reliable guide to real-world differences” at these capability levels. Don’t switch your production traffic based on a press release.
Instead: take 50–100 real examples from your own product, run them through Opus 5.5, GPT-6 Sol, and Luna, and score the outputs the way your users would. What you’re looking for is the cheapest model that passes your quality bar, not the one with the best benchmark chart. One afternoon of evals can permanently cut your bill.
Step 4: Mind the migration details
A couple of gotchas worth knowing. Opus 5.5’s API has breaking changes relative to Opus 5 — some integrations will need updates, and thinking mode can no longer be disabled. On the OpenAI side, GPT-5.5-instant retires October 14, so if you depend on it, plan the migration now rather than during a fire drill. And note that Opus 5.5 is text-output only, so multimodal pipelines still need other models.
The Catch: What Nobody’s Putting in the Headline
I want to be fair here, because the coverage of this launch wave has been pretty breathless. A few honest caveats:
Per-token price isn’t your real cost. What you pay per task = (tokens used) × (price per token) × (how often you retry or correct). A cheaper model that needs twice as many tokens — or that you have to babysit — can cost more than the expensive one. Opus 5.5 claims to use fewer tokens per task, which compounds the savings; Luna’s value depends heavily on your routing discipline. Measure cost per completed task, not cost per token.
The 40% figure is measured at default settings. Anthropic’s headline “40% less to run than Opus 5” is measured at each model’s default effort level, which is medium on Opus 5.5 and high on Opus 5. If you run everything at max effort, your mileage will differ. The saving is real, but the exact number depends on your configuration.
Nobody has independently verified the benchmarks. As of writing, every performance claim is the vendor’s own. That doesn’t mean they’re false — both labs have decent track records — but I’d treat benchmark tables as marketing until third parties replicate them. MarkTechPost’s writeup has the full table if you want to dig in, and PYMNTS summarized the cost angle well.
Prices can go back up. “Promotional pricing” is doing a lot of work in these announcements. GPT-6 Sol’s price is half of GPT-5.6 Sol’s promotional price. Build your unit economics on today’s prices with a margin of safety — don’t price your product assuming $0.10 input tokens are a permanent law of nature.
Frequently Asked Questions
Is Claude Opus 5.5 better than Claude Fable 5.1?
Anthropic says Opus 5.5 performs at Fable 5.1’s level “on most work,” while Fable remains the top-tier model for creative and narrative tasks. Think of Opus 5.5 as the practical flagship — near-top performance at meaningfully lower cost — with Fable reserved for the most demanding creative work. Honestly, for most business workloads (coding, agents, knowledge work), Opus 5.5 is the one to evaluate.
Should I switch from GPT-5.6 models to GPT-6 Sol or Luna?
If you’re on GPT-5.6 Sol, the move to GPT-6 Sol is close to a no-brainer on paper: half the price with better quality. For GPT-5.6-instant-style high-volume work, Luna at $0.10/$0.50 is dramatically cheaper. But run your own evals first — “better on the vendor’s benchmark” and “better on my users’ tasks” are different things. And note GPT-5.5-instant retires October 14, 2026, so some migration is coming regardless.
What’s the cheapest capable model for high-volume AI workloads?
Right now, GPT-6 Luna ($0.10 in / $0.50 out per million tokens) is the cheapest frontier-lab model for routine tasks, and OpenAI claims high-effort Luna matches GPT-5.6 Sol at roughly 1/100th the cost. For coding-heavy agent work where cache reads dominate, Opus 5.5’s $0.20 cache reads are the number to beat. The honest answer is it depends on your workload mix — which is why Step 3 of the playbook (your own evals) matters more than any chart.
Does the “40% cheaper” claim apply to my usage?
Maybe. Anthropic’s 40% figure is for “typical workloads” at default settings, driven by the combination of lower list prices, 60% cheaper cache reads, and fewer tokens per task. If your workload is heavy on long-context agent loops (cache reads), you’ll likely see close to that. If you’re mostly doing short single-turn calls, the saving will be closer to the 20% list-price cut. Measure your own mix.
Will AI model prices keep falling?
The trend is clearly downward — open-weight models keep improving, inference keeps getting more efficient, and the labs are competing on price for the first time in a serious way. I’d bet on continued declines for commodity-grade intelligence. But the frontier will stay expensive: the hardest reasoning, the longest contexts, and the most capable agents still command premium pricing. Plan for a widening gap, not a uniform drop.
The Bottom Line
Two labs launched four models in 90 minutes and declared a price war. For founders, the takeaway isn’t “which model won” — it’s that intelligence just got cheaper across the board, and the winners will be the teams that act on it fastest.
Here’s my one-week action list: audit your token spend, route easy tasks to cheap models, turn on prompt caching, and re-run your evals on Opus 5.5, Sol, and Luna. That afternoon of work could cut your AI bill by a third or more — and it might unlock the features you’ve been pricing out of existence.
And if you’re still in the planning stage, this is genuinely good news: the cost floor for building AI products keeps dropping. Our list of profitable AI startup ideas for 2026 got a little more profitable yesterday, and the toolkit every founder should be using just got cheaper to run. The price war is on — might as well be the one who benefits from it. If you’d rather have professionals handle the build, our web agency AISquadX designs and ships AI-powered products for startups.
