Tool Reviews

Claude Opus 5.5, GPT-6 Sol and Luna arrive — and spark a fresh pricing battle

Anthropic launched Claude Opus 5.5, cutting Opus pricing 20% to $4/$20 with cache reads down 60%, while OpenAI quickly followed with GPT-6 Sol and Luna at half their predecessors' cost — Luna at just $0.10/$0.50 per million tokens, undercutting nearly the whole market. The price battle targets mid-tier models, pressuring rivals like Grok 4.7 and leaving GPT-5.6 Terra unattractive. Opus 5.5 promises clearer communication and Fable-level intelligence, though its max-thinking mode once burned through token limits without delivering output.

Key Takeaways

  • A wave of launches in one week: xAI shipped Grok 4.7, Xiaomi released MiMo v2.6 Flash/Pro, Anthropic launched Claude Opus 5.5, and OpenAI followed within an hour with GPT-6 Sol and GPT-6 Luna.
  • Pricing, not benchmarks, is the real story: GPT-6 Luna costs just $0.10/$0.50 per million input/output tokens — half its predecessor's price — while Opus 5.5 got its first meaningful cut ($5/$25 → $4/$20, with cache reads down 60% to $0.20/M).
  • GPT-5.6 is effectively obsolete: With a scheduled 25% price increase in November, GPT-6 is effectively half the cost, leaving little reason to pick GPT-5.6 Terra over GPT-6 Sol.
  • "Max" thinking can backfire: Opus 5.5 at max reasoning became the first model to fail the author's pelican-on-a-bicycle SVG test — it over-thought the prompt, exhausted its 128K output limit, and produced nothing, twice ($2.56 and ~20 minutes each).
  • Good week to re-benchmark your stack: GPT-6 Sol and Opus 5.5 are now defaults in Codex and Claude Code; high-volume app teams should evaluate GPT-6 Luna, while those needing max reasoning may prefer Fable 5.1 for now.

Developers evaluating AI models woke up to a crowded field this week: one day after xAI shipped Grok 4.7 and Xiaomi released MiMo v2.6 Flash and Pro, Anthropic launched Claude Opus 5.5, and OpenAI followed roughly an hour later with two new models, GPT-6 Sol and GPT-6 Luna. Early impressions suggest the biggest story is not benchmark scores but pricing — the new GPT-6 models cost half as much as their GPT-5.6 predecessors, and Opus 5.5 saw its first meaningful price cut in the Opus line. Here is what the shake-up means for anyone building on these APIs.

GPT-6 pricing undercuts the entire market

GPT-5.6 Luna was already considered excellent value for application development. GPT-6 Luna comes in at half that price again: $0.10 per million input tokens, $0.01 per million cached input tokens, and $0.50 per million output tokens. GPT-6 Sol received a similar reduction relative to GPT-5.6 Sol.

The current pricing landscape:

ModelInputCached inputOutput
GPT-6 Luna$0.10/M$0.01/M$0.50/M
GPT-5.6 Luna$0.20/M$0.02/M$1.20/M
Grok 4.7$2/M$0.50/M$6/M
GPT-6 Sol$2/M$0.20/M$10/M
GPT-5.6 Terra$2/M$0.20/M$12/M
Claude Opus 5.5$4/M$0.20/M$20/M
GPT-5.6 Sol$4/M$0.40/M$20/M
Claude Fable 5.1$10/M$0.25/M$50/M
GPT-6 Astra$10/M$1/M$50/M

There is an additional wrinkle: GPT-5.6 has a scheduled 25% price increase for November, meaning GPT-6 is effectively half the price of what the older models will cost going forward. With GPT-5.6 Terra now priced at or above GPT-6 Sol's level, there is little reason left to pick Terra.

The competitive pressure is visible across vendors. Grok 4.7 had positioned itself at $2/$6, less than half of GPT-5.6 Sol's price — but GPT-6 Sol now matches it on input and comes far closer on output. At $0.10/$0.50, GPT-6 Luna is among the cheapest models OpenAI has ever released, undercut by only two far smaller models: GPT-4.1 Nano ($0.10/$0.40, April 2025) and GPT-5 Nano ($0.05/$0.40, August 2025).

At the top end, GPT-6 Astra and Claude Fable 5.1 both sit at $10/$50, so the price war so far affects the tier below the flagship models.

Opus 5.5 gets cheaper and (hopefully) more pleasant to talk to

Claude Opus models have long kept a single price: from Opus 4.5 through Opus 5, every version charged $5 per million input tokens and $25 per million output tokens. Opus 5.5 breaks that pattern with a 20% cut to $4/$20.

The bigger drop is in cache reads, which fell 60% to $0.20 per million tokens. That matters a great deal for long agentic conversations, where 90% or more of input tokens are typically processed at cached rates.

Anthropic's Thariq Shihipar described the release as driven by user feedback on Opus's communication style: the model "communicates clearly," is cheaper per token than Opus 5.0, carries "the intelligence of Fable 5.1," is token-efficient, and works across every effort level. Anthropic also claims improvements in Blender, which will be worth testing.

One open question: Anthropic says Sonnet 5.5 and Haiku 5.5 are coming soon. Current Haiku 4.5 costs $1/$5 — ten times the price of GPT-6 Luna — so whether Haiku can stay competitive at the low end is one of the more interesting upcoming tests.

When "max" thinking is too much thinking

For years, the author's informal benchmark — prompting a model to "Generate an SVG of a pelican riding a bicycle" — has produced usable results. Claude Opus 5.5 at the "max" thinking level became the first model to fail it outright, twice.

The model began by calling the request "a classic test request" and then reasoned in extraordinary depth: calculating shin length at "roughly 95.2," planning leg paths from hip to knee to ankle, placing a fish in the basket for the pelican to eye, and checking chainring teeth and layer ordering. Then it stopped. Claude models, including Opus 5.5, have a 128,000 maximum output token limit, and the model burned through it while still reasoning — never delivering the SVG.

The two failed attempts cost $2.56 each and took nearly 20 minutes apiece. The practical takeaway: "max" may be effectively unusable if it can over-think a trivial prompt to the point of producing nothing. By contrast, Claude Fable 5.1 on "max" did not over-think and produced the best pelican yet from any Anthropic model.

Comparing across vendors via bicycle-riding pelicans may not measure much, but the test remains useful for comparing reasoning levels within a single model family. One stylistic pattern stood out: the GPT-5.6 family chose bolder, brighter colors, while the GPT-6 family's output is noticeably more muted. GPT-6 Astra on max still produced the best pelican overall.

What this means for your stack

For developers, the near-simultaneous releases make this a good week to re-benchmark default models. Early real-world usage suggests the new options are already production-ready: GPT-6 Sol and Claude Opus 5.5 are now serving as default models in Codex and Claude Code respectively, and the Datasette Agent demo at agent.datasette.io has been upgraded to GPT-6 Luna, where it handles both SQL queries and building HTML and JavaScript for Datasette Apps quickly and competently.

What to watch next: independent benchmarks on the new models, whether Anthropic's Sonnet 5.5 and Haiku 5.5 can close the pricing gap at the low end, and whether GPT-5.6's November price increase pushes the remaining holdouts to migrate.

Who should pay attention / who should wait: Teams building high-volume applications should look hard at GPT-6 Luna, whose pricing reshapes the budget math for cheap-tier workloads. Developers who found previous Opus versions verbose or expensive have concrete reasons to try Opus 5.5. Anyone relying on "max" reasoning effort should test carefully before upgrading — and may be better served by Fable 5.1 at that tier for now.

Comments (0)

  1. No comments yet. Be the first to share what worked for you.

Leave a comment

Comments are reviewed before they appear. Your email address is not published.