Two AI companies shipped competing models hours apart on the same day, and both of them cut prices. That has never happened before. Usually a launch is a benchmark story: new model, new leaderboard, same pricing. This week it was a pricing story with benchmarks attached, and if you pay an AI bill every month, the numbers matter more than the leaderboards.
Here is the setup. On September 22, 2026, Anthropic released Claude Opus 5.5, the first model in its new 5.5 family. Within hours, OpenAI released GPT-6 Sol and GPT-6 Luna, mid-tier and budget models that slot under the flagship GPT-6 Astra. Both launches dominated the Hacker News front page at the same time. Both companies framed the release around cost. And the two price sheets, laid side by side, tell you exactly where each lab thinks the market is going.
One disclosure before the math: these are launch-day numbers from both companies' official pages and independent coverage. Neither model has been out long enough for serious long-term testing by anyone. What follows is a comparison of what each lab is offering and what it should cost you, not a verdict from months of production use. Treat it as a buying framework, not a review.
The price sheets, side by side
Anthropic prices Opus 5.5 at $4 per million input tokens and $20 per million output tokens. That is a 20 percent cut from Opus 5, which shipped in July at $5 and $25. The bigger cut is on cache reads: $0.20 per million tokens, down 60 percent from $0.50. Anthropic says cache reads make up the majority of agentic and coding work costs, which is why that number matters more than the headline input price.
OpenAI prices GPT-6 Sol at $2 per million input and $10 per million output, with cached input at $0.20. GPT-6 Luna lands at $0.10 and $0.50, with cached input at a single cent. OpenAI says these are permanent prices, not a launch promotion, and that they are roughly half of what the GPT-5.6 models they replace cost.
Put the two headliners next to each other:
- Opus 5.5: $4 input / $20 output / $0.20 cached input per million tokens
- GPT-6 Sol: $2 input / $10 output / $0.20 cached input per million tokens
- GPT-6 Luna: $0.10 input / $0.50 output / $0.01 cached input per million tokens
- Old Opus 5, for reference: $5 input / $25 output / $0.50 cached input
Sol is exactly half of Opus 5.5 on fresh tokens. The cached input price is identical at $0.20. And Luna is not really competing with Opus at all. It costs 40 times less on both input and output, which puts it in the territory hosted open-weight models used to own.
What the math says about your actual bill
Headline prices mislead, because almost nobody pays the uncached input rate. Agent workloads re-read the same large context dozens of times per task, so the realistic cost picture mixes cache reads, fresh tokens, and output. Here is what the numbers work out to for a typical agentic workload: one million input tokens per month with 90 percent served from cache, plus 100,000 output tokens.
- Opus 5.5: about $2.58 per million-input-equivalent workload
- GPT-6 Sol: about $1.38, so roughly 47 percent cheaper
- GPT-6 Luna: about $0.07, essentially rounding error
- Opus 5 (old pricing): $3.45, which shows how much Anthropic's own cut matters
Scale that to a heavier month, say 10 million input tokens and a million output tokens, and the gap becomes real money: roughly $26 on Opus 5.5, $14 on Sol, and $0.69 on Luna. If your workload is cache-heavy, the Opus 5.5 versus Sol gap narrows a bit because their cache read prices are identical; the difference comes almost entirely from fresh input and output tokens.
Two pricing footnotes worth knowing before you commit. OpenAI charges double the input and cache rates, with 1.5x output, for the entire request when a prompt exceeds 272K input tokens on Sol and Luna. Anthropic offers a fast mode on Opus 5.5 at up to 2.5x speed for $8 and $40 per million tokens. Both models ship with roughly a million-token context windows (1M on Opus 5.5, 1.05M on Sol and Luna), so the surcharge thresholds, not the windows, are the practical limit to watch.
Benchmarks: close, thin, and mostly unverified
Here is the honest part. Nobody has run these two models head-to-head on independent harnesses yet, and the early picture is fragmentary.
Anthropic's claim for Opus 5.5 is that it matches Claude Fable 5.1, its previous frontier model, "on most tasks" while costing 40 percent less to run than Opus 5. The company also says the model generates output more than 30 percent faster and uses fewer tokens per task, which is where part of that 40 percent total-cost claim comes from. If accurate, the token-efficiency point matters as much as the price cut: a model that spends fewer tokens at a slightly higher rate can still win on bill.
OpenAI's GPT-6 Sol, meanwhile, has one early independent data point. Artificial Analysis ran it through their Coding Agent Index in OpenAI's Codex harness, where Sol at max effort scored 57, up 2 points from GPT-5.6 Sol, with gains on Terminal-Bench 4.0 (43 percent versus 37) and SWE-Atlas-QnA (58 versus 54). A 2-point gain on a 100-point index is an improvement, not a generational leap.
For context on the prior generation, the Opus 5 versus GPT-5.6 Sol split was genuinely divided: Opus 5 led SWE-bench Pro by a wide margin while GPT-5.6 Sol led on some terminal and real-world-change benchmarks. Nothing so far suggests the new generation collapses that split into a clean winner. The New Stack put it plainly: OpenAI halved prices to beat Anthropic on cost, but nobody has run the two new models against each other yet.
The reasonable read one day in: performance between Opus 5.5 and Sol is close enough that price, not benchmarks, should drive the decision for most workloads. That is exactly what both pricing teams are betting on.
The subscription chess move nobody is talking about
The API price war is only half the story. Both companies also moved on subscription terms the same day, and this is where the competition gets interesting.
Anthropic removed the five-hour usage caps anxiety in a different way: five-hour limits on Pro, Max, Team, and seat-based Enterprise plans increase by 20 percent, and because the model is cheaper to run, Anthropic says those limits effectively stretch 25 percent further. Subscribers also get a rate-limit reset they can save and use whenever they choose, which softens the single worst part of Claude's subscription experience, hitting a wall mid-task.
OpenAI countered from below. Paid ChatGPT plans get both Sol and Luna in Work and Codex immediately, and free users get Luna on the desktop app. That last part is the aggressive move: OpenAI is putting a current-generation model in the hands of every free user on day one, at $0.10 and $0.50 per million tokens of serving cost. That is a subscriber-acquisition play funded by a 58 percent cut in Luna's output price.
- If you are on Claude subscriptions: the 20 percent limit increase plus the saved reset is a genuine quality-of-life upgrade, independent of model quality.
- If you are on ChatGPT free tier: you now have a frontier-family model available without paying, which was not true yesterday.
- If you are choosing a team plan: model availability in the tools you already use (Codex, Claude Code) may matter more than either price sheet.
So which one should you actually pick?
Here is the decision framework I would use, and one I expect to hold even as benchmarks fill in over the coming weeks.
Pick GPT-6 Sol when cost per task dominates. Bulk summarization, classification, high-volume agent loops, background processing. At exactly half the price of Opus 5.5 with identical cache read pricing, Sol is the rational default for anything you run at scale. If your current setup cannot tell the difference between models in output quality, the 2x price difference is free money.
Pick Luna when the task is clerical. OpenAI's own framing is that Luna handles summarizing, extracting, and quick questions. At $0.69 per month for a 10-million-token workload, it costs less than the coffee you drink while reviewing its output. If you are still running cheap older models for preprocessing, Luna probably replaces them at similar or lower cost with better quality.
Pick Opus 5.5 when output quality per task is the bottleneck. Hard refactors, long agentic coding sessions where a failure costs an hour of retry loops, work where token efficiency compounds. Anthropic's 40 percent total-cost claim means the real gap with Sol is smaller than the sticker prices suggest, especially for cache-heavy coding workloads. And SWE-bench Pro style results have historically favored Claude's line by large margins; until independent numbers say otherwise, that is the safer bet for repository-scale work.
Watch the 272K threshold. If your prompts run long on the OpenAI side, the 2x surcharge above 272K input tokens can erase Sol's price advantage entirely. Opus 5.5 has no equivalent cliff in its published pricing.
The meta-point is bigger than either model. For two years, frontier-adjacent quality meant frontier pricing, and the choice was between expensive and very expensive. Both of these launches push the same direction from opposite sides: Anthropic cut its own flagship line by 20 percent and its costs by 40 percent, and OpenAI halved its mid-tier outright. The price of "good enough to ship" intelligence fell by roughly half in a single day. Whatever you were paying in August, your bill has no excuse to look the same in October.
The takeaway
If you only remember one thing: benchmark differences this small do not justify paying 2x, so start from Sol's price sheet and only pay Opus money when you have evidence you need it. And if you run free-tier tools, Luna's arrival means the floor just moved. Check what model your default tools are serving this week; there is a decent chance it quietly got better and cheaper.
What are you running these days, and did either launch change your pick? I am genuinely curious whether the cache-heavy math holds up in other people's workloads, because that $0.20 cache read price on both sides is doing a lot of work in this comparison.
I write about AI, developer tools, and the engineering decisions behind them every week. Subscribe, it is free, and it makes sure the next price war lands in your feed instead of passing you by.
Sources: Anthropic's Opus 5.5 announcement, OpenAI's API pricing page and GPT-6 Sol/Luna model docs, The New Stack's launch coverage, Artificial Analysis's GPT-6 benchmarking notes, and independent coverage from unite.ai, Cryptobriefing, and Benzinga. Pricing and benchmark figures are as reported on September 22 to 23, 2026.
Top comments (0)