In the last few days multiple large language model updates arrived. After earlier launches of Grok 4.7 and MiMo v2.6 Flash/Pro, Anthropic released Claude Opus 5.5 and about an hour later OpenAI published GPT-6 Sol and GPT-6 Luna. It will take more testing to form a complete picture, but early impressions highlight substantial price changes and some behavioral differences.
Price moves — notable reductions
The most striking effect is on pricing. GPT-6 Luna and GPT-6 Sol are materially cheaper than their GPT-5.6 predecessors. GPT-6 Luna is listed at $0.10 per million input tokens, $0.01 per million cached input, and $0.50 per million output tokens. By comparison, GPT-5.6 Luna was priced at $0.20 / $0.02 / $1.20 per million (input / cached input / output).
Other examples noted in the coverage include:
- Grok 4.7: $2 / $0.50 / $6 per million (input / cached input / output)
- GPT-6 Sol: $2 / $0.20 / $10 per million
- GPT-5.6 Terra: $2 / $0.20 / $12 per million
- Claude Opus 5.5: $4 / $0.20 / $20 per million
- GPT-5.6 Sol: $4 / $0.40 / $20 per million
- Claude Fable 5.1: $10 / $0.25 / $50 per million
- GPT-6 Astra: $10 / $1 / $50 per million
Context matters: GPT-5.6 models had a scheduled 25% price increase set for November, so GPT-6 pricing is attractive even relative to promotional 5.6 rates. With GPT-5.6 Terra priced the same as GPT-6 Sol, any remaining price-based reasons to use Terra effectively disappeared.
GPT-6 Luna’s $0.10 / $0.50 pricing places it among the cheapest OpenAI models ever released, surpassed only by earlier, much smaller Nano-class models.
Anthropic’s Opus 5.5 and its price cut
Anthropic positions Opus 5.5 as addressing the main user complaints about Opus’s communication style. Thariq Shihipar (quoted in the coverage) described Opus 5.5 as the product of user feedback: clearer in communication, cheaper per token than Opus 5.0, possessing Fable 5.1-level intelligence, token-efficient, and usable across effort levels. It is also intended to be better at Blender-style tasks.
Previous Opus versions (4.5 through 5) had been priced at $5 per million input and $25 per million output. Opus 5.5 reduces those rates by 20% to $4 per million input and $20 per million output. The cache-read price fell by 60%, which is significant for longer, agentic conversations where a very large share of input tokens can be billed at cached rates.
Opus 5.5’s new price matched GPT-5.6 Sol’s prior price point — but that parity changed once OpenAI halved Sol prices.
Anthropic also announced Sonnet 5.5 and Haiku 5.5 are coming soon. It remains to be seen whether Haiku can reclaim price competitiveness in the low tier, given the present Haiku 4.5 price of $1 / $5 while GPT-6 Luna is $0.10 / $0.50.
A performance hiccup: Opus 5.5 over-thinks at "max"
In practical testing, Opus 5.5 exhibited a problematic behavior: on the simple prompt “Generate an SVG of a pelican riding a bicycle,” Claude Opus 5.5 at the "max" thinking level failed to return a response. The model produced a long, detailed internal planning trace — describing geometry, anatomy, coordinates, layer order, viewBox dimensions and so on — and continued reasoning until it hit the model’s maximum output token limit of 128,000 while still in the midst of planning.
A second attempt produced the same outcome. The author notes that each failure cost roughly $2.56 and took nearly 20 minutes. This suggests that the "max" setting can engage in excessive internal computation that ultimately prevents a usable response for some prompts. By contrast, Claude Fable 5.1 at "max" did not over-think and produced what the author called the best pelican from any Anthropic model.
Visual comparisons
The author rendered pelicans under GPT-6 Luna and GPT-6 Sol and compared those to renders from Opus 5.5, Opus 5, Fable 5.1 and Sonnet 5 in a comparison grid. Subjectively, the 5.6 family used bolder, brighter colors while the 6 family tended to be more muted; the author preferred the GPT-6 Astra render at maximum settings.
Practical adoption
Following these releases, the author switched default models to GPT-6 Sol and Claude Opus 5.5 for Codex and Claude Code workflows. They also upgraded the Datasette Agent demo (agent.datasette.io) to use GPT-6 Luna, reporting that Luna performs quickly and competently for SQL queries and for generating HTML and JavaScript for Datasette Apps.
Summary
The new model launches have intensified price competition, particularly in the low- and mid-tier model segments, and shifted the economics users consider. Performance differences remain important: Opus 5.5 shows improvements in communication and token efficiency but can be hampered by excessive internal reasoning at high effort settings. Anthropic and OpenAI both plan further releases, so users will soon have more data to decide which models best fit their applications.
