Yesterday felt like watching two tech giants simultaneously drop their prices and call each other’s bluff. Grok 4.7 arrived with aggressive pricing, Anthropic released Claude Opus 5.5, and then OpenAI dropped GPT-6 Sol and GPT-6 Luna within an hour. The pricing landscape shifted so dramatically that some older models instantly became obsolete.
Let me be direct: GPT-6 Luna at $0.10 per million input tokens and $0.50 per million output tokens is absurdly cheap. It’s now one of the cheapest models OpenAI has ever released, beaten only by GPT-4.1 Nano and GPT-5 Nano. But here’s what matters more than the headline number: this represents a systematic reduction across the entire pricing tier.
GPT-5.6 Luna was already my favorite model for building applications because it combined performance with cost efficiency. GPT-6 Luna is half the price again. GPT-6 Sol saw similar reductions. Meanwhile, Grok 4.7 started at $2/$6 and is now essentially price-matched to GPT-6 Sol on input tokens. This isn’t a sale. This is competitive pressure materializing in real time.
What Changed in Claude Opus 5.5
Anthropics’s new Opus 5.5 addresses the biggest complaints about the previous version: communication clarity and token efficiency. The pricing reflects this positioning: $4 per million input tokens and $20 per million output tokens, a 20% reduction from Opus 5.0. More importantly, cache read prices dropped 60%, which is transformative for agentic systems where 90% of input tokens hit cached pricing.
I tested Opus 5.5 at maximum reasoning on my standard SVG pelican test, and something unexpected happened: it failed. The model over-thought the problem so severely that it hit the 128,000 token output limit mid-reasoning without generating the SVG. This cost $2.56 and took nearly 20 minutes. I tried again and got the same result.
This failure mode concerns me. If “max” reasoning breaks on a trivial prompt, I don’t trust it with complex work. Fable 5.1 on max reasoning, by contrast, performed flawlessly and produced the best pelican I’ve seen from any Anthropic model. This suggests the reasoning implementation in Opus 5.5 needs refinement.
The Real Implications
The price war matters because it’s collapsing the tier structure between models. GPT-6 Astra and Claude Fable 5.1 remain at $10/$50, but everything below is getting compressed. Current Haiku pricing at $1/$5 is now 10x more expensive than GPT-6 Luna. If Anthropic doesn’t aggressively cut Haiku pricing when 5.5 ships, they lose the value segment entirely.
For developers building applications, this changes the calculus. I’ve already migrated my Datasette Agent demo to GPT-6 Luna, and performance is strong on SQL generation and HTML/JavaScript building. The cost reduction is significant enough that I’m questioning whether to maintain multiple model implementations at all. Why optimize for a $1/$5 model when you can use a $0.10/$0.50 alternative?
The pelican comparisons show stylistic differences: the GPT-5.6 family uses bolder, brighter colors while GPT-6 uses muted tones. But from a functional standpoint, both families produce usable output on my test. GPT-6 Luna is slightly weaker than GPT-6 Sol, but the 80% price reduction seems worth the capability trade-off for most production use cases.
Looking Forward
I’m now using GPT-6 Sol and Claude Opus 5.5 as default models in my applications. GPT-6 Sol hits the sweet spot between capability and cost at a tier that’s now equivalent to Opus 5.5 pricing (which was already competitive before OpenAI halved their Sol prices).
Sonnet 5.5 and Haiku 5.5 are coming soon from Anthropic, and I’m watching carefully to see if Haiku can reclaim price competitiveness. The current gap between Haiku 4.5 ($1/$5) and GPT-6 Luna is unsustainable for Anthropic’s positioning. Either Haiku drops dramatically, or Anthropic concedes the cost-sensitive segment entirely.
What we’re witnessing isn’t just price competition. It’s a fundamental restructuring of which models justify inclusion in your application stack, and whether the capability differences between tiers are large enough to justify the cost gaps anymore.