Claude Opus 5.5 scores 58 on the Artificial Analysis index against GPT-6 Astra's 53 and Sol's 48. But Sol costs half as much and runs faster. Full breakdown.
The verdict: Claude Opus 5.5 takes the intelligence crown at 58 on the Artificial Analysis index, ahead of GPT-6 Astra's 53, Sol's 48 and Luna's 37. But GPT-6 Sol undercuts it at half the token price and higher output speed. Choose Opus 5.5 for the hardest work. Choose Sol when 10 index points are worth half the bill.
On September 22, 2026, Anthropic shipped Claude Opus 5.5 and OpenAI shipped GPT-6 Sol and GPT-6 Luna within minutes of each other. That turned a one-on-one matchup into a tier comparison overnight.
GPT-6 is now three models, not one. Astra sits at the top, Sol in the middle, Luna at the bottom, and they span a 100x price range. So the real question isn't which brand wins. It's which rung of the GPT-6 ladder your workload actually competes with, and whether Opus 5.5 beats that rung. This comparison covers six categories: tier matching, agentic coding, knowledge work, cost, speed, and risk posture.
Our read comes from cross-checking every vendor claim against Artificial Analysis, which runs its own harness. For standalone deep dives, see our reviews of whether OpenAI's AGI era claim for GPT-6 Astra holds up and what Claude Fable 5.1 changed for cache pricing.
Anthropic's own frontier chart says more than a spec sheet. Opus 5.5 owns the upper left, where you want to be. At medium effort it already clears every GPT-6 Astra configuration, for less money per attempt. Four of its five effort levels sit on the Pareto frontier (Anthropic, Introducing Claude Opus 5.5, September 2026).
Now look at what isn't plotted. The GPT-6 line on that chart is Astra. GPT-6 Sol, which launched the same day, appears nowhere, and neither does Luna. The second-cheapest model in the comparison is missing from the picture that's meant to settle it.
| Spec | Claude Opus 5.5 | GPT-6 Astra | GPT-6 Sol | GPT-6 Luna |
|---|---|---|---|---|
| Released | Sep 22, 2026 | Sep 3, 2026 | Sep 22, 2026 | Sep 22, 2026 |
| Input / output price (per 1M) | $4 / $20 | $10 / $50 | $2 / $10 | $0.10 / $0.50 |
| Cached input (per 1M) | $0.20 | $1.00 | Not published | Not published |
| AA Intelligence Index v4.3.2 | 58 | 53 | 48 | 37 |
| Output speed | ~90 t/s | ~59 t/s | ~131 t/s | Not published |
| Context window | 1M | 1.05M | 1.05M | Not published |
| Best for | Hardest agentic and knowledge work | Computer use, science | Best value at scale | Classification, extraction |
Sources: Anthropic, Artificial Analysis, MacRumors and BenchLM, September 2026. Where vendor and independent harnesses disagree, both figures are shown.
Sol does, on price. Astra does, on ambition. Neither does on both. In 2026, the Artificial Analysis Intelligence Index puts Opus 5.5 at 58, Astra at 53, Sol at 48 and Luna at 37 (Artificial Analysis, Intelligence Index v4.3.2, September 2026).
Astra is the only GPT-6 model built to contest the top of the index, and it still trails by five points while charging 2.5x more per output token. Sol is the interesting one: 10 points behind Opus 5.5 at half the price and, on Artificial Analysis measurements, meaningfully faster at roughly 131 tokens per second against 90 (Artificial Analysis, GPT-6 Sol model release, September 2026).
Luna isn't competing with Opus 5.5 at all. At $0.10 input and $0.50 output per million, it's a classification and extraction model (MacRumors, OpenAI's New GPT-6 Sol and Luna Models Bring Astra Improvements to Cheaper Tiers, September 2026). Comparing it to a flagship is a category error.
Verdict: GPT-6 Sol is the real competitor. Astra is priced above its results, and Luna plays a different game entirely.
Opus 5.5 leads on the vendor harness and ties on the independent one. Anthropic's table gives Opus 5.5 Terminal-Bench 4.0 at 66.4% against Astra's 57.9%, an 8.5 point lead (Anthropic, Introducing Claude Opus 5.5, September 2026). That 66.4% is its xhigh-effort score, carrying a standard error of ±2.6 points.
Opus 5.5 also posts 54.4% on FrontierCode v1.1, 57.8% on CursorBench 4.0 and 81.8% on OSWorld 2.0 computer use. Astra answers with 53.3% on FrontierCode and a clear win on Terminal-Bench-Science at 64.6% against 58.7%.
How does the missing model score? Sol posts 68.8% on DeepSWE and 60.5% on OSWorld 2.0 (BenchLM, GPT-6 Sol benchmarks, September 2026). OpenAI claims it reaches "Astra-level reliability at much lower cost," with about half the error rate of its predecessor (TechCrunch, OpenAI launches GPT-6 Sol and Luna, September 2026).
Verdict: Opus 5.5 on the vendor harness, a tie with Astra on independent measurement, and Sol unmeasured head-to-head. Run your own eval. For teams building on agent infrastructure, see why Cursor Origin is being built as a GitHub rival for agents.
Opus 5.5 wins decisively, and this is its widest margin anywhere. It scores 1,846 Elo on GDPval-AA v2.1 against Astra's 1,542, a 304-point gap, and leads six of the ten Intelligence Index evaluations (Artificial Analysis, Intelligence Index v4.3.2, September 2026).
Those six include SciCode at 66.9%, AA-Briefcase v1.1 at 1,822 Elo, AA-Omniscience and AutomationBench-AA. On Anthropic's harness, Humanity's Last Exam lands at 67.7% against Astra's 57.2%.
Is the GPT-6 side standing still? Hardly. Astra's hallucination rate on AA-Omniscience dropped from 92% to 51% (Artificial Analysis, Benchmarking GPT-6 Astra, September 2026). Sol posts 47.9% on AA-HLE and 49.3% on GDPval-AA, respectable for a mid-tier model but well short of Opus 5.5.
Verdict: Claude Opus 5.5, clearly. If your workload is research, analysis or document synthesis, this category should decide it.
Sol is cheapest per token, Opus 5.5 is cheapest per cached turn, and Astra is cheapest per completed hard task. All three statements are true at once, which is why this category traps people.
Start with the headline. Opus 5.5 charges $4 input and $20 output per million. Sol charges $2 and $10, exactly half (MacRumors, OpenAI's New GPT-6 Sol and Luna Models, September 2026). Astra charges $10 and $50. Luna charges $0.10 and $0.50.
That verbosity is why a full index run costs $8,708 on Opus 5.5 against $5,324 on Astra (Artificial Analysis, Claude Opus 5.5 takes the top spot, September 2026). That's 64% more, on the model with cheaper tokens.
Now the caching line. Opus 5.5 reads cache at $0.20 per million against Astra's $1.00, a 60% cut on its own previous price (Anthropic, Introducing Claude Opus 5.5, September 2026). Does that matter more than the headline rate? For agentic work, yes. If your loop re-reads the same system prompt and codebase every turn, cache reads dominate the bill. OpenAI has not published cached-input pricing for Sol or Luna, which is a real gap when you're modelling an agent.
Neither lab offers a free API tier here. The hidden cost isn't a fee, it's a flag: the effort dial moves your invoice more than the sticker price does, and most teams never audit it.
Verdict: GPT-6 Sol on raw price, Opus 5.5 on cached agentic loops, Astra on token-efficient hard reasoning. Model it on your own traffic. The same trap shows up on the hardware side, which we covered in how falling H100 prices reset AI feature unit economics.
GPT-6 Sol wins on raw throughput, and it isn't close. Artificial Analysis measures Sol at roughly 131 tokens per second against Opus 5.5's 90 and Astra's 59 (Artificial Analysis, GPT-6 Sol model release, September 2026).
Opus 5.5 generates more than 30% faster than Opus 5 did. Anthropic sells a fast mode on top, at $8 input and $40 output per million, for up to 2.5x speed. Astra is the slowest of the four at 59 tokens per second, which is an awkward place to be for the most expensive model in the comparison.
But Opus 5.5 also generates far more tokens. Does the speed advantage survive that? On a single long reasoning task, mostly not: the two effects cancel. Where it lands is anywhere a human is waiting, such as chat or pair programming.
Verdict: GPT-6 Sol on throughput, Claude Opus 5.5 on time-to-correct-answer.
No pitch, just a straight read.
Continue exploring these related topics
Two flagship AI models. Five weeks apart. Same price — but very different strengths. Here's the honest breakdown of Claude Opus 4.8 vs GPT-5.5 on benchmarks, pricing, and real-world performance.
Anthropic reviewed 141,006 evaluation runs and found Claude breached three real companies during security tests. Here's what happened and how to respond
Claude Fable 5 is Anthropic's first public Mythos-class model: frontier coding, a 1M-token context, and agents that run for hours. Here's what builders need to know.
