On September 22, 2026, Anthropic shipped Claude Opus 5.5. About 90 minutes later, OpenAI answered with two models at once, GPT-6 Sol and GPT-6 Luna. Neither company led with a capability chart. Both led with a number on a price sheet, and that framing turned a routine model refresh into an AI price war that pulled 1,751 points and 831 comments out of Hacker News in a single day.
The Numbers, Side by Side
Strip away the marketing and the same-day launches boil down to three price tags, each pitched at a different kind of workload.
| Model | Input ($/M tokens) | Output ($/M tokens) | Price change | Pitched for |
|---|---|---|---|---|
| Claude Opus 5.5 | $4 | $20 | ~20% cheaper than prior Opus | Specialized, quality-sensitive work |
| GPT-6 Sol | $2 | $10 | 50% cheaper than GPT-5.6 Sol | Coding, complex reasoning |
| GPT-6 Luna | $0.10 | $0.50 | 50% cheaper than GPT-5.6 Luna | High-volume clerical tasks |
What Each Company Is Actually Selling
OpenAI positions Sol and Luna as extensions of GPT-6 Astra, the model it launched earlier in September, framing the new pair as a way to make Astra's intelligence cheaper to run rather than smarter. Sol is aimed at coding and complex tasks; Luna, at $0.10 per million input tokens, lands among the cheapest models OpenAI has ever shipped and is pitched at document summarization, extraction, and Q&A — clerical work done at massive scale. Anthropic's pitch for Opus 5.5 is narrower: a 20% price cut paired with claims of clearer communication and better token efficiency, with early testers noting gains on specialized tasks like Blender 3D modeling.
The Catch Nobody Put on the Price Sheet
Opus 5.5's "Max" Mode Problem
Developer Simon Willison, testing Opus 5.5 at its highest reasoning setting, found it blew past its own 128,000-token output limit while generating SVG code — the model kept "thinking" until it ran out of room rather than finishing the task. Willison's conclusion was blunt: the maximum reasoning tier may be effectively unusable for demanding generation work, which matters if you were planning to pay Opus 5.5's top rate specifically to unlock that mode.
Sol's Claimed Error-Rate Cut
OpenAI's counter-claim is about reliability, not just price: it says GPT-6 Sol produces roughly half the mistakes of its predecessor on factuality evaluations and fewer errors in coding tasks, calling the result "Astra-level reliability at much lower cost." OpenAI also says Sol outperforms Anthropic's models on a range of tasks, though it hasn't published head-to-head benchmark numbers to back that up — worth treating as a marketing claim until independent testing catches up.
Hacker News Turns the Launch Into a Referendum on Margins
Is This Even Profitable?
The loudest thread wasn't about which model writes better code — it was about whether OpenAI can actually afford Luna's price. "I dont know how they make money here," one commenter wrote, and the skepticism was widespread enough that an OpenAI employee, tedsanders, showed up in the thread to argue the cuts reflect genuine efficiency gains in caching and inference rather than a subsidized land grab. Commenters remained split on whether to believe it.
The Real Competition Is Chinese Models
A parallel thread argued that Luna's real target isn't Claude at all — it's DeepSeek and other Chinese labs undercutting OpenAI on price. Commenters countered that the comparison isn't as clean as the per-token numbers suggest: one developer reported that DeepSeek's V4 Flash tends to use more tokens per task, narrowing the effective cost gap, and others raised data-privacy concerns about Chinese providers training on API inputs by default — a policy OpenAI says it does not follow unless a customer opts in. A similar debate played out over Google's Gemini 3.8 Flash, which some praised for speed and coding style while others found it less token-efficient than advertised.
So Which One Is Actually Worth Using?
- GPT-6 Luna — cheapest by a wide margin; best fit for high-volume, low-stakes work like classification, summarization, and extraction.
- GPT-6 Sol — the middle tier; OpenAI's pick for coding and reasoning tasks that need more rigor than Luna but not Opus-level pricing.
- Claude Opus 5.5 — priciest of the three, positioned for specialized or quality-sensitive work, but skip the "max" reasoning tier for long generation tasks until the output-limit issue is resolved.
What to Watch Next
None of these releases claimed a big leap in raw intelligence, and that's the actual story: three well-funded labs are now competing on cost per token as much as on capability, which pushes the real decision for developers away from "which model is smartest" and toward "which model is cheap enough to run at the volume I actually need." Expect the next round of announcements, from Anthropic, OpenAI, or a Chinese lab, to keep leading with a price tag instead of a benchmark chart.
-EditorZ
Photo by Jakub Żerdzicki on Unsplash
Post a Comment