A New Price War: Anthropic and OpenAI Slash Their Flagship Rates

The release cadence keeps accelerating. Yesterday brought Grok 4.7 and MiMo v2.6; today Anthropic shipped Claude Opus 5.5, and roughly an hour later OpenAI put out both GPT-6 Sol and GPT-6 Luna. Getting a solid read on all of them will take time, but here are the early impressions.

GPT-6 costs half of GPT-5.6

GPT-6 Sol and Luna are priced at half their GPT-5.6 counterparts. The previous-generation Luna was already a favorite for building applications, pairing excellent performance with a remarkably low price. Somehow GPT-6 Luna is half that again: $0.10 per million input tokens and $0.50 for output. That makes it one of the cheapest models OpenAI has ever released, undercut only by the far weaker GPT-4.1 Nano and GPT-5 Nano.

One caveat: GPT-5.6 has a 25% price hike scheduled for November, so GPT-6 is effectively half the promotional pricing of its predecessors. And with GPT-5.6 Terra now costing the same as GPT-6 Sol, any lingering reason to use Terra has simply evaporated.

It is hard to overstate how aggressive this pricing is. Grok 4.7 launched at \(2/\)6, less than half the price of GPT-5.6 Sol — but it now matches GPT-6 Sol on input and sits close on output.

Opus 5.5 got cheaper too

Claude followed suit. Opus versions from 4.5 through 5 all held a single price of \(5/\)25 per million tokens; Opus 5.5 cuts that by 20%, to \(4/\)20. The drop in cache-read pricing is even steeper — 60% — which matters for long agentic conversations, where more than 90% of input tokens are billed at cached rates. According to its makers, the release addresses the main complaints about Opus's communication style: it speaks more clearly, uses tokens more sparingly, and is better at Blender.

Opus 5.5's price landed level with GPT-5.6 Sol — but that was before OpenAI halved Sol's rate. The top tier, GPT-6 Astra and Claude Fable 5.1, still sits at \(10/\)50; the price war is playing out one rung below. Anthropic promises Sonnet 5.5 and Haiku 5.5 soon, and whether Haiku can reclaim its low-end edge is an open question: today's Haiku 4.5 costs \(1/\)5, ten times GPT-6 Luna.

When "max" breaks itself

A curious failure came with the "draw an SVG of a pelican riding a bicycle" test. For the first time ever, Opus 5.5 on the "max" reasoning level returned nothing at all. It called the prompt "a classic test request" and then thought very, very hard: verifying the shin length (about 95.2 — "checks out"), plotting the leg path from hip to pedal, mulling over the fish poking out of the basket and whether to give the pelican a happier expression. All that deliberation ran straight into the 128,000-token output ceiling, and the model stopped without ever producing the image. A second attempt ended the same way. Each failure cost the author $2.56 and took nearly 20 minutes. The lesson is blunt: if "max" can self-lock on a silly drawing, trusting it with real work is risky. Fable 5.1 on "max," by contrast, didn't overthink and delivered the best pelican yet from any Anthropic model.

Source: Simon Willison