Two frontier flagships landed within a month of each other: Claude Opus 5 on July 24, 2026, and GPT-5.6 Sol, announced June 26. They cost the same to feed and different amounts to listen to, they sit two points apart on the leading independent intelligence ranking, and one of them has a billing rule that can nearly double your invoice without changing a line of your code. Here is the comparison with the numbers from our own live models database.
Quick answer
Claude Opus 5 currently ranks first — 61 on the Artificial Analysis Intelligence Index against 59 for GPT-5.6 Sol — and it is cheaper on output ($25 vs $30 per million tokens), which works out to a blended $10.00 vs $11.25 on typical traffic. GPT-5.6 Sol’s advantage is raw context (1.05M vs 1M tokens) and the OpenAI ecosystem. The decisive detail most comparisons miss: on requests above 272K input tokens, GPT-5.6 Sol bills at 2× input and 1.5× output — so long-context work costs far more than the headline price.
Key facts
- Intelligence: Opus 5 ranks #1 of 170 models at 61 on the Artificial Analysis index; GPT-5.6 Sol scores 59
- Price: identical $5 input; output diverges — $25 (Opus 5) vs $30 (Sol)
- Blended cost (3:1 input:output): $10.00 vs $11.25 per million tokens
- Context: 1M (Opus 5) vs 1.05M (Sol); both cap output at 128K tokens
- Hidden cliff: Sol requests over 272K input tokens bill at 2× input / 1.5× output ≈ $10/$45
- Fast mode: Opus 5 offers an API-only Fast tier at $10/$50
Claude Opus 5 vs GPT-5.6 Sol at a glance
| Claude Opus 5 | GPT-5.6 Sol | |
|---|---|---|
| Developer | Anthropic | OpenAI |
| Released | July 24, 2026 | Announced June 26, 2026 |
| Input price / 1M | $5.00 | $5.00 |
| Output price / 1M | $25.00 | $30.00 |
| Blended (3:1) | $10.00 | $11.25 |
| Context window | 1,000,000 | 1,050,000 |
| Max output | 128K | 128K |
| Intelligence index | 61 (#1 of 170) | 59 |
| Long-context billing | Flat | 2× input / 1.5× output above 272K |
| Alternate tier | Fast mode $10/$50 (API only) | — |
The 272K cliff: the number that decides real bills
This is the part worth reading twice. GPT-5.6 Sol advertises a 1.05-million-token context — the largest of the two — but any request carrying more than 272,000 input tokens is billed at double the input rate and one and a half times the output rate for the entire request. In practice that turns $5/$30 into roughly $10/$45, and pushes the blended cost from $11.25 to about $18.75 per million tokens.
The irony is sharp: the capability you would buy Sol for — very long context — is precisely the mode where its price advantage disappears. Opus 5 bills flat across its full million-token window. If your workload regularly ships large codebases, long document sets or big retrieval payloads, model both scenarios in our AI API cost calculator before committing.
Capability: a narrow lead, honestly stated
On the Artificial Analysis Intelligence Index — the most-cited independent aggregate — Opus 5 sits first of 170 models at 61, with GPT-5.6 Sol at 59 and Anthropic’s own Fable 5 between them at 60. Two points is a real but modest gap, and it is not the kind of margin that should override ecosystem fit or existing integrations.
Anthropic’s own benchmarking claims more dramatic movement: Opus 5 more than doubles Opus 4.8’s score on Frontier-Bench v0.1 and scores roughly three times the next-best model on ARC-AGI 3, a reasoning benchmark designed to resist memorisation. Treat vendor benchmarks as directional — but the independent index agrees on the ordering.
Where each one fits
- Agentic coding and long-horizon tasks → Opus 5. It is built around end-to-end software work, code review and coordinating parallel subagents, and it is cheaper per output token — which matters because agents generate output relentlessly.
- Very large single-shot context → GPT-5.6 Sol has the bigger window, but budget for the 272K cliff; above that threshold Opus 5 is both cheaper and flat-rated.
- Existing OpenAI or Azure infrastructure → Sol, for the integration you already have.
- Cost-sensitive high volume → neither. Both are premium tiers, and our price-performance index shows efficiency-tier models delivering many times the capability per dollar on routine work.
Convly’s take
Opus 5 wins this comparison on the merits available today — first on the independent index, cheaper on output, and flat-rated at long context — but the more useful conclusion is that the two are close enough that neither should be your only model. The pattern that actually saves money in 2026 is routing: a frontier model for the hard 10% of tasks, an efficiency-tier model for the rest. Teams that pick one flagship and send everything through it are typically paying five to ten times more than the work requires — and the gap between Opus 5 and Sol is a rounding error next to that decision.
Frequently asked questions
Is Claude Opus 5 better than GPT-5.6 Sol?
On the Artificial Analysis Intelligence Index, yes — Opus 5 ranks #1 of 170 models at 61 versus 59 for GPT-5.6 Sol. It is also cheaper on output tokens. The gap is real but narrow, so ecosystem fit often matters more.
How much does Claude Opus 5 cost?
$5 per million input tokens and $25 per million output tokens — unchanged from previous Opus pricing. An API-only Fast mode is billed at $10/$50.
How much does GPT-5.6 Sol cost?
$5 input and $30 output per million tokens for standard requests. Requests above 272K input tokens are billed at 2× input and 1.5× output for the whole request — roughly $10/$45.
Which has the bigger context window?
GPT-5.6 Sol, at 1.05M tokens versus 1M for Opus 5. Both cap output at 128K tokens. Note that Sol’s long-context pricing changes above 272K input tokens, while Opus 5 bills flat.
Which is better for coding?
Opus 5 is positioned around end-to-end software tasks, code review and multi-agent coordination, and its lower output price suits agentic workloads that generate a lot of tokens. Sol is strong on command-line and multi-step coding too — see our AI coding agents comparison.
More comparisons: Claude vs ChatGPT · Kimi K3 vs Claude Opus 4.8 · the full models database · self-hosting vs API calculator.
