Opus 5.5, GPT-6 Sol, and GPT-6 Luna just got cheaper. The effort setting decides if you notice

Anthropic shipped Claude Opus 5.5 yesterday: the first release in its new Opus 5.5 line, priced 40% below Opus 5 at default settings, and matching Claude Fable 5.1 on most work by Anthropic's own account. OpenAI answered within a day with GPT-6 Sol and GPT-6 Luna, both priced 50% below their GPT-5.6 predecessors. GPT-6 Astra, still OpenAI's flagship, kept its price where this site already reported it: $10 input and $50 output per million tokens. Two labs, two real price cuts, the same week.

The more useful read isn't the price sheet. It's a chart both companies published on the same announcement page as the discount, and it undercuts the headline before anyone else gets the chance to.

The chart that comes with the discount

Opus 5.5's list price is a genuine cut: input tokens went from $5 to $4 per million, output from $25 to $20, and cache reads, which Anthropic says make up most agentic and coding costs, from $0.50 to $0.20. A few scrolls down the same page, Anthropic's own Terminal-Bench 4.0 result plots cost per attempt on a log scale running from roughly $2 to $20, labeled low, medium, high, xhigh, and max: the same model, at different effort settings, spanning close to a 10x range in what one task costs to run. A 20% cut on the input rate barely registers next to that.

Read past the benchmark row you're being sold

OpenAI's launch page makes a related point from the competitive angle. On AutomationBench, GPT-6 Sol at its highest effort setting completes a task for $0.27. GPT-6 Astra at its lowest effort setting costs 3.9 times that on the same benchmark. Claude Opus 5 at max effort costs 11.1 times that. None of those three numbers involve this week's discounts at all, they're what effort setting alone does to a bill, across three different models.

The footnote on that table is worth reading too: Claude Fable 5.1's listed cost "understates its actual cost, as it omits the cost of the Opus 5 fallbacks, which occurred on ~40% of tasks." That's a competitor describing a competitor's number, so read it as marketing, not neutral. But the mechanism cuts both ways: any agent that quietly falls back to a pricier model once a task gets hard enough doesn't show that cost on the headline benchmark row either. If your own workflow has a fallback path, its real cost lives in your logs, not on anyone's launch page.

The number a price cut can't fix for you

Both labs are showing you the same thing without quite meaning to: cost per task moves with effort setting more than it moves with list price, by a wide margin. That's exactly why checking cost per task instead of cost per token matters more this week than most. A Time MCP Server call or a Fetch MCP Server request stays narrow and deterministic no matter which model sits behind it, so a price cut on the base rate lands close to full value. A broad, long-running agent like Modernization Agent, reading an entire codebase before proposing anything, lives at a high effort tier by design, and that's where the effort curve, not the price cut, decides the bill.

Three questions before you switch anything

Before moving a workload to Opus 5.5, GPT-6 Sol, or GPT-6 Luna because the price dropped, check three things in your own usage, not the announcement: what effort setting does this workload run at in production right now? What was the average cost per completed task last month, at that setting, not the demo? And does the published discount apply to the effort tier you actually run, or only to the cheapest one, the one the headline number is quoting?

Anthropic and OpenAI both published the chart that answers this. It sits a few scrolls below the price table, and it matters more than the number above it.