New

Opus 5, GPT-5.6, Kimi K3: three launches, three rankings that contradict each other

Claude Opus 5 (July 24, 2026), GPT-5.6 (July 9) and Kimi K3 (July 27) shipped weeks apart. Pricing, benchmarks, and why the WebDev Arena crown changed hands in three days.

  • Comparison
  • Tooling
Published

Quick answer

Three frontier models shipped weeks apart. GPT-5.6 on July 9, 2026 , Claude Opus 5 on the 24th , and Kimi K3 from Moonshot AI on the 27th, just three days later . On this site, Opus 5 is the new default model for Claude Max and the most capable one available on Claude Pro , which is reason enough for an article. But what makes this week interesting isn't "who wins": it's that all three labs published rankings that contradict each other from one day to the next, sometimes from one table to another in the same document.

This article compares the technical specs, shows where Opus 5 genuinely leads the pack, and where the numbers making the rounds deserve a more careful read than the headline attached to them.

The three spec sheets

On paper, all three models play in the same price bracket, with comparable but not identical approaches to reasoning "effort."

ModelInput ($/MTok)Output ($/MTok)ContextMax outputDefault effort
Claude Opus 5$5$251M tokens128k tokenshigh (5 levels: low to max)
GPT-5.6 Sol$5$301.05M tokens128k tokensnot publicly documented
Kimi K3$3 (cache miss) / $0.30 (cache hit)$151M tokens (1,048,576)131,072 by default, up to 1,048,576max (3 levels: low, high, max)

Opus 5's pricing hasn't moved compared to Opus 4.8: the same $5 / $25 per million tokens. Kimi K3 is a case apart: it's an open-weight model, a MoE architecture with 2.8 trillion total parameters of which 104 billion are activated per token, released under a proprietary Moonshot license (not a modified MIT license, contrary to what's sometimes claimed: past $20 million in annual revenue on "Model as a Service" usage, a separate commercial agreement is required). We'll come back to it in detail in a dedicated article.

The one index that (almost) everyone agrees on

Artificial Analysis publishes an Intelligence Index that aggregates several independent evaluations. On the July 28, 2026 snapshot, Claude Opus 5 (adaptive thinking, max effort) tops the chart with a score of 61, ahead of Claude Fable 5 (60) and GPT-5.6 Sol (59). And Opus 5 is also cheaper there: $3.85 per million tokens at blended pricing versus $4.35 for Sol.

The nuance to keep in mind: Opus 5 is noticeably slower in throughput, 54.8 tokens per second against 76.0 for Sol, even though its time to first token is better (65 seconds versus 143). First on the index, yes. The fastest at producing text, no.

Another nuance: these numbers move depending on the measurement date. Kimi K3's tech report, frozen on July 23 (the day before Opus 5's launch), shows a different snapshot: GPT-5.6 Sol at 58.9, Fable 5 at 59.9, Opus 4.8 at 55.7, Kimi K3 at 57.1, fourth out of 580 models. It's not a contradiction, it's a snapshot taken on a given day: Opus 5 didn't exist yet when Moonshot stopped its measurement run.

The WebDev Arena crown changed hands in three days

This is the most telling story of this launch week. WebDev Arena is a human-voted leaderboard that ranks the quality of code produced by a model on web development tasks.

Kimi K3's tech report claims the top spot: 1,678 Elo, snapshot dated July 23, 2026, presented as "the first open model to top this leaderboard." The document was published on July 27, the very day the model launched.

Except the live leaderboard snapshot, on that same July 27 date, no longer has Kimi K3 in the lead. Claude Opus 5 (the max variant) moved ahead with 1,725 points, Kimi K3 (max) follows at 1,682, then Opus 5 (high variant) at 1,670, Fable 5 at 1,629, and GPT-5.6 Sol (xhigh) at 1,623. So the tech report that proudly announces the top open spot shows, on the very day of its own publication, a ranking that's already outdated.

This isn't a mistake on Moonshot's part: its evaluations were frozen before Opus 5 even launched. It's simply proof that on this kind of leaderboard, a number has a short shelf life.

The GPQA trap: these numbers don't come from Anthropic

If you're looking for a GPQA Diamond or AIME score for Claude Opus 5, you won't find one in Anthropic's own publications. Opus 5's system card contains no occurrence of "GPQA" or "AIME." The only GPQA Diamond numbers circulating for Claude models (Fable 5 at 92.6%, Opus 4.8 at 92%) actually come from the comparison table OpenAI published in its own GPT-5.6 launch post, not from a measurement Anthropic ran itself. If an article cites "Claude scores XX% on GPQA," always check which lab actually ran the measurement: Anthropic simply doesn't publish this benchmark.

swebench.com doesn't know Opus 5 or GPT-5.6

The official SWE-bench Verified leaderboard, maintained by the SWE-bench team itself, still has no row for Opus 5 or GPT-5.6 as of July 28, 2026. Its most recent entries date back to February 2026; the best score there is 76.80% for "Claude 4.5 Opus" at high reasoning.

The 96.0% SWE-bench Verified figure often cited for Opus 5 comes from section 8.2 of Anthropic's system card, measured internally. It's a self-reported score, not an entry validated by the official leaderboard: harness, conditions, and date all differ, so a direct comparison with the 76.80% above doesn't hold up.

The internal inconsistencies in OpenAI's own post

Even staying within the single GPT-5.6 launch document, two passages contradict each other between the text and the numbers table next to it.

On Agents' Last Exam, the article's text announces "a new high of 53.6" and a 13.1-point gap over Claude Fable 5. The table gives 52.7% for GPT-5.6 Sol and 40.5% for Fable 5, a gap of 12.2 points, not 13.1.

On BrowseComp, the text claims "state-of-the-art results at 92.2%" for Sol. The table attributes that 92.2% to a distinct variant, "Sol Ultra," and gives 90.4% for Sol alone. Two details that don't change the overall conclusion (GPT-5.6 remains very strong on both tests) but that are a good reminder to always read the table, not just the text next to it.

What actually changes in Claude Code

1

Opus 5 is now your default model on Max

On Claude Max, Opus 5 replaces Opus 4.8 as the default model. On Claude Pro, it's the most capable model available. You don't need to configure anything to benefit from it.

2

Default effort is `high`

Opus 5 supports five effort levels (low, medium, high, xhigh, max), and Claude Code uses high by default. If your longest tasks were stalling on the previous model, now's the time to try xhigh or max.

3

A request with no `thinking` field now runs with adaptive thinking

This is a breaking change if you call the API directly. On Opus 4.8, a request without a thinking field ran without thinking enabled. On Opus 5, it runs with adaptive thinking. If your scripts never send this field explicitly, check the behavior before migrating to production.

4

`thinking: disabled` with `xhigh` or `max` effort returns a 400 error

Second breaking change: combining thinking: {"type": "disabled"} with xhigh or max effort is no longer accepted on Opus 5, unlike on Opus 4.8. You either need to raise thinking back to low or high, or remove the thinking field.

{
"model": "claude-opus-5",
"effort": "xhigh",
"thinking": { "type": "disabled" }
}

This request returned a valid result on Opus 4.8. On Opus 5, it returns a 400 error: at this effort level, thinking can no longer be disabled.

And where does Kimi K3 fit in?

Kimi K3 doesn't run in Claude Code: it's an open-weight model that you deploy via vLLM, SGLang, or its own TokenSpeed engine, or use through Moonshot's OpenAI- and Anthropic-compatible API. It only appears in this article because it's at the center of the WebDev Arena ranking story. Its full profile (license, architecture, detailed benchmarks) is the subject of a separate article.

FAQ

Is Opus 5 really better than GPT-5.6 Sol? On the Artificial Analysis Intelligence Index from July 28, 2026, yes: 61 versus 59, and at a lower blended cost. But Sol produces text almost 1.4 times faster. Neither is "better" in absolute terms: it depends on what you're measuring.

Is Opus 5 really number 1 on WebDev Arena? As of the July 27, 2026 snapshot, yes, with 1,725 Elo. But with only 686 votes against several thousand for its direct competitors, this top spot isn't statistically solid yet.

Why can't I find an official GPQA score for Claude Opus 5? Because Anthropic doesn't publish it. The GPQA numbers attached to Claude models come from OpenAI's comparison tables, not from Anthropic itself.

Is Opus 5's 96% SWE-bench Verified independently verified? No. It's a self-reported figure from Anthropic's own system card. The official swebench.com leaderboard has neither Opus 5 nor GPT-5.6 as of July 28, 2026.

Should I change my API scripts after Opus 5's launch? Check two things: if you never send a thinking field, your request now activates adaptive thinking by default; and if you combine thinking: disabled with xhigh or max effort, you'll get a 400 error.

Next steps