Andy Garcia · freelance dev
FRENES
Back to the blog
AI & Pricing

Opus 5.5 and GPT-6 Sol, 90 Minutes Apart: The AI Price War

September 23, 2026 5 min read

I build a product on Claude. So on Tuesday, September 22, I did what a lot of CTOs did that afternoon: I opened two launch pages side by side and tried to work out what my bill would look like next month.

Anthropic released Claude Opus 5.5 first. According to TechCrunch, OpenAI followed about 90 minutes later with GPT-6 Sol and GPT-6 Luna. Both announcements led with the same promise, more capability for less money, and both came with a kind of chart I did not see on launch pages a year ago: accuracy plotted against cost per task, on a log scale.

That second detail is the real story. But first, the numbers everyone quoted.

The sticker prices

Opus 5.5 costs $4 per million input tokens and $20 per million output tokens. That is 20 percent less than Opus 5, which was $5 and $25. Cache writes drop from $6.25 to $5. Cache reads drop from $0.50 to $0.20, a 60 percent cut. There is also a fast mode, up to 2.5 times faster, at $8 and $40.

GPT-6 Sol costs $2 input and $10 output, exactly half of GPT-5.6 Sol's $4 and $20. GPT-6 Luna, the small one, costs $0.10 and $0.50, down from $0.20 and $1.20. OpenAI describes the cut as 50 percent against the promotional pricing of the 5.6 generation.

Read like that, Sol is half the price of Opus 5.5 on every token. That is the headline most people ran with, and for my workload it is close to meaningless.

The line that matters is the cache

Anthropic says it plainly in its own post: cache reads "make up the majority of agentic and coding work costs." That matches what I see. An agent that works for an hour rereads the same system prompt, the same files and the same conversation hundreds of times. The fresh tokens are a thin layer on top of a thick, reused base.

So look at the cache. Opus 5.5 reads cached tokens at $0.20 per million. OpenAI says its improved prompt caching for GPT-6 gives a discount of 90 percent on cached input reads, which on Sol's $2 input price also comes to $0.20 per million.

On the part of the bill that dominates agent work, the two models cost the same.

What remains is how many tokens each model burns to finish a job. Which brings us back to those charts.

Cost per task is now the unit

Anthropic's claim for Opus 5.5 is not really about price per token. It says the model is cheaper per token and uses fewer tokens per task, which "nets out to a 40% drop in costs" against Opus 5 on typical workloads. The customer quotes on the page repeat the pattern. GitHub says that in VS Code it solved more terminal tasks than Opus 5 in less than half the steps. Optiver reports matching Opus 5's quality in about half the turns, time and output tokens, cutting the cost of that workload by 40 to 50 percent. Spotify talks about completing the same tasks cheaper and faster.

OpenAI makes the same move from the other side. Its headline comparison is on AutomationBench, a test of business workflows across apps: GPT-6 Sol at its "xhigh" effort outperforms Claude Opus 5 at max effort "at just 9% of Opus 5's cost per task." On OpenAI's factuality evaluation, Luna at higher effort matches GPT-5.6 Sol at about a hundredth of its cost.

Both companies now sell you a cost per completed task. That is the right unit. It is also a unit you cannot read off a price list, because it depends on the task.

Each launch compares itself to the model the other was replacing

Here is the detail that made me smile. Anthropic's benchmark table compares Opus 5.5 with GPT-6 Astra, OpenAI's flagship, and with GPT-5.6 Sol. OpenAI's tables compare GPT-6 Sol with Claude Opus 5 and Fable 5.1.

Nobody compared against the new thing, because 90 minutes earlier the new thing did not exist. Anthropic measured itself against the Sol that OpenAI was about to replace. OpenAI measured itself against the Opus that Anthropic had just replaced.

That is not a scandal. It is how launches work. But it means that on the evening of September 22, there was no official comparison of the two models most teams were actually choosing between.

The one join I could make

There is one benchmark on both pages with the same reference score. On AutomationBench, both companies list Claude Opus 5 at 26.9 percent. That gives a common anchor.

On Anthropic's page, Opus 5.5 scores 40.0 percent and GPT-6 Astra 41.4 percent. On OpenAI's page, GPT-6 Sol at xhigh effort scores 33.2 percent, at $0.27 per task.

I would not bet a contract on that comparison. Anthropic notes that its Opus 5.5 run comes from Zapier's own evaluation during early access, with safeguard interventions counted as failures, and OpenAI cites a specific version of the benchmark. Different runs, different settings. But it is the only like-for-like line available, and it says something the headlines did not: the cheaper model is not doing the same work.

What I am actually going to do

Three things, none of them glamorous.

First, I am going to stop comparing prices per token. I will rerun our own evaluation set on each model and record cost per completed task, including retries and the tokens spent on tool calls. The only benchmark that sets my bill is my own traffic.

Second, I am going to treat vendor tables as two halves of a join I have to do myself. If a vendor chart leaves out the competitor I am actually considering, that is information, not an oversight to forgive.

Third, I am going to plan for this to keep happening. Anthropic says Sonnet 5.5 and Haiku 5.5 will follow in the coming weeks. OpenAI shipped GPT-6 Astra earlier this month and two cheaper siblings less than three weeks later. The price of a given level of capability is now falling in steps of weeks, not years.

For a product team, that changes one architectural rule. The model is no longer a decision you make once. It is a parameter you revisit every month, and the cost of revisiting it is the quality of your evaluation harness. Teams with a good one will ride the price war. Teams without one will read launch pages, like I did on Tuesday afternoon.

Sources

A project like this one?

I design and deploy products like this. Let's talk.

Let's talk