Opus 5.5 vs GPT-6 Sol and Luna

September 23, 2026 · Episode Links & Takeaways

Author

AI Daily Brief
September 23, 2026

MAIN STORY

MAIN STORY

Opus 5.5 vs. GPT-6 Sol and Luna

For the first time in memory, OpenAI and Anthropic shipped new models on the same day. OpenAI's GPT-6 Sol and Luna push further on cost efficiency at every level of the intelligence stack, while early reactions suggest Opus 5.5 is a return to the kind of acclaim Anthropic hasn't enjoyed since Opus 4.6. The same-day release invites a "who won" framing, but the more useful question is where each model fits in the rotation.

Claude Opus 5.5
Fable 5.1-level performance at roughly 40% lower running cost.
Opus 5.5 beats both Opus 5 and Fable 5.1 on every benchmark Anthropic published, including a jump on TerminalBench 4.0 from Fable 5.1's 55.8% to 66.4%, and trails GPT-6 Astra only on Automation Bench and Terminal Bench Science. List pricing drops 20% to $4/$20 per million input/output tokens, with Anthropic saying efficiency gains push the real savings to about 40%; the model also averages 30% faster output than Opus 5, and five-hour usage limits were raised on Pro, Max and Team plans alongside a one-time rate limit reset. It is the first model released since the Pacing the Frontier note, tested pre-release by external evaluators including METR, and Anthropic's Sam Bowman says it is sufficiently safer than its predecessors that releasing it "more likely than not reduces risks related to misalignment."

GPT-6 Sol and Luna
"Half the price per token and even less per task."
If Opus 5.5 was partly about efficiency, Sol and Luna were entirely about it: much of Astra's strength in faster, cheaper models, with API prices 50% below the previous generation's. OpenAI frames lower cost as a new way of working, with more room to iterate, which extends the long-running split between Claude for hands-off, long-running tasks and GPT models in Codex for interactive work. Tellingly, OpenAI dropped traditional benchmark tables entirely in favor of performance-vs-cost charts, and says the 90th percentile of its researchers use $7,000 of tokens per day.

The Independent Benchmarks
Sol and Luna win on efficiency; Opus 5.5 wins outright.
On Artificial Analysis's new V3 index, GPT-6 Luna scores roughly even with GPT-5.6 Luna at 37 and Sol edges past its predecessor, but both get there at lower cost, with significantly reduced hallucination. Zapier found GPT-6 Sol scored 33.2% on real workflows versus 28.77% for GPT-5.6 Sol at about half the price, and its advice was simple: upgrade. Opus 5.5 was a different story, jumping five points to 58 on the Intelligence Index, ahead of the previous leaders Fable 5.1 and GPT-6 Astra at 53, while also getting a 20% price cut.

Box
63% fewer tokens, 42% less verbose, 30% faster.
Box ran Opus 5.5 on complex enterprise knowledge work involving unstructured data and found major efficiency gains over Opus 5, along with accuracy gains of 39% on financial services tasks, 65% on cloud cost analysis and technology use cases, and 17% and 15% on consumer products and clinical diagnostics respectively.

Visuals, 3D and Code-Generated Worlds
"Just as good at building 3D scenes as Astra."
Much of the demo discourse was visual, and Opus 5.5 appears to match or beat GPT-6 Astra on the exact Blender and 3D work that defined Astra's launch. Demos included single-prompt claymations, a code-generated Golden Gate Bridge, a 1906 Market Street reconstruction with no image generation, and paintings rendered pixel by pixel in Python. Once coding ability can one-shot a 30-second marketing animation or a site redesign trailer, these showpieces start opening real use cases.

Writing and Personality
"We fixed the writing."
Writing is where near-universal agreement exists that Opus 5.5 is strictly better. Anthropic says it communicates more naturally, puts the most important information first and follows the writing rules it's given. Every called it the most readable prose it has seen from an Anthropic or OpenAI model, but said the bigger upgrade is the feel: it takes feedback without a fight and builds on material instead of handing it back tidier.

The Safeguards Problem
Overzealous guardrails make it a non-starter for some.
Because Opus 5.5 is comparable to Mythos 5.1 in biology and cybersecurity, it ships with Fable-style safeguards. Users doing legal work report more safety refusals and poor effort budgeting versus Fable 5.1, and life science companies that fell back on Opus when Fable refused biology requests have lost that option. It isn't widespread, but where it hits, it makes the model effectively unusable.

The Cost Curve
Jevons paradox, applied to agents.
Aaron Levie argues that every drop in AI cost dramatically expands what agents can be deployed against, driving broader diffusion into the economy. Epoch AI backs this with data showing that, at a given level of performance, AI costs have fallen around 47% per quarter since 2023: four times faster than DNA sequencing, six times faster than compute and 18 times faster than lithium batteries.

Who Won?
"Sol feels like an S-class iPhone release."
On buzz, Opus 5.5 took the day, with a real "Claude is back" wave of people returning from Codex. But the strategies differ: OpenAI is aggressively building a cost-efficient model slate for an era of multi-model architectures, while Anthropic set out to reclaim momentum, and did. Every's verdict: Sol is the fast, affordable daily driver; Opus 5.5 is worth paying more for ambitious coding and visual work.

Personality Is UX
Even ruthless early adopters have personality preferences.
Most of the Opus 5.5 excitement isn't about new capabilities but about how pleasant it is to use. The GPT-4o deprecation backlash wasn't just a normie phenomenon: AI is a tool that approximates a personality, and how enjoyable that personality is affects how much people get out of it.

The Battle for the Stack
The fight is for each slot, not one crown.
Even Opus 5.5's biggest fans aren't saying to use it for everything. What matters now is understanding a complete set of needs and how models and harnesses combine to fill them across performance, cost and speed.

Models and Harnesses
Models can no longer be separated from harnesses.
For users invested in Codex, a better Opus doesn't matter if OpenAI's options are close enough. The real test in a week is whether anyone actually switched, or whether everyone tested and returned to the ecosystem holding their context, tools and rules.

Muse
Traction where users don't know or care about the model.
Meta's Muse may be the first widely adopted AI product whose users are indifferent to the underlying model, raising the question of whether the product era, where models disappear into the system, is finally arriving.

Pacing the Frontier
"Hot take, this is what pacing looks like."
Four major releases in two days prompted jokes, but none of them was Astra- or Fable-tier, and that's the point. Models that fix specific problems and deliver incremental gains with major cost efficiencies are a good place to spend time, and not a sign that adoption will slow. After years of raw capability jumps doing the work, the labs are being pushed to think in product terms and in terms of human and system realities.