GPT-5.6 vs Claude Opus 5 vs Kimi K3 vs Everyone Else: Which Model Should You Actually Build On?

Six frontier-class models shipped in three weeks. Leaderboard position stopped being the decision, here's the framework that actually holds up.

The question "which AI model is best" stopped being useful sometime in July. Six frontier-or-close-to-it models shipped inside three weeks, OpenAI's GPT-5.6 family, Anthropic's Claude Opus 5, Google's Gemini 3.6 Flash, Moonshot's Kimi K3, Alibaba's Qwen3.8-Max, and DeepSeek's V4-Flash-0731, and every one of them tops a leaderboard somewhere. Leaderboard position isn't the decision anymore. The decision is: what's the job, and what does it cost to do that job a million times a month.

OpenAI split itself into three models on purpose

GPT-5.6 didn't ship as one model, it shipped as three: Sol, Terra, and Luna, priced at $5/$30, $2.50/$15, and $1/$6 per million input/output tokens at launch. That's not a tiering gimmick, it's an admission that "one model for everything" was never the right shape for how people actually use these things. Sol gets a Max Reasoning Effort mode and an Ultra Mode that runs sub-agents in parallel, built for the hardest problems you'll hand it. Luna is built to be cheap enough to run on every classification call and every lightweight agent step without thinking about the bill. Most teams building anything agentic will end up routing between at least two of the three, not picking one and calling it done.

Claude Opus 5 didn't get more expensive to get better

Anthropic's Opus 5, released July 24, holds the same $5/$25 per million token price as Opus 4.8, which is the actual headline. It scores 96.0% on SWE-bench Verified against Fable 5's 95.0%, though Fable 5 still edges it out on the harder SWE-bench Pro (80.0% vs 79.2%). The practical read: Opus 5 is now the default model on Claude Pro at no price change, so if you're already paying for Claude, you got a real capability bump for free. If you're deciding between Opus 5 and Fable 5 for a coding-heavy build, the benchmark gap is close enough that your actual codebase and task type will matter more than the leaderboard delta.

The open-weight models stopped being the budget option

This is the part that should actually change how you think about model selection. Kimi K3 shipped July 16 with a 2.8 trillion parameter MoE architecture and a 1M-token context window. DeepSeek's V4-Flash-0731 landed July 31 under a genuinely permissive MIT license, scoring 82.7 on Terminal Bench 2.1, up from 72.1 on its own preview, at $0.14/$0.28 per million tokens, the cheapest production-grade model in this entire group. (Note added September 8, 2026: this article's stated publishedAt is July 29, 2026, but the passage above references a July 31 release, so at least one referenced fact was not actually available on the stated publish date. We don't have a record of when this specific passage was written or last edited, so we're flagging the anachronism itself rather than asserting it was added in a later revision.) Qwen3.8-Max claims 86.1 on OSWorld-Verified, ahead of both GPT-5.6 Sol Max (83.2) and Fable 5 (85.0) on computer-use tasks, at $2/$6 per million tokens, under a third of Opus 5's combined rate. (Our directory's Qwen family page is a lineage/background page for the Qwen family generally, not a dedicated review of the Max variant specifically; treat it as background context on the family rather than a Qwen3.8-Max-specific evaluation.) That Qwen number is Alibaba's own benchmark and independent replication is still pending, worth flagging before you build a decision around it. But even discounting that one claim, the pattern holds: open-weight models from Chinese labs are now competitive on raw capability and dramatically cheaper, not just "good enough for a budget tier."

Context window stopped being a differentiator

Six months ago, a 1M-token context window was a selling point. Now Gemini 3.6 Flash, Kimi K3, Claude Opus 5, GLM-5.2, and DeepSeek V4-Flash all publish 1M-token windows as table stakes. If a vendor is still leading with context size in their pitch, that's a signal they don't have anything more current to lead with. The axes that actually separate these models now are cost per task at real volume, whether the weights are open and what the license actually permits commercially, and which benchmark protocol a claimed score came from, OSWorld 2.0 and OSWorld-Verified are different tests measuring different things, and a lot of coverage conflates them.

A framework that survives the next model wave

Stop asking "which model is best" and ask three questions instead. First: is this a high-volume, low-complexity job, classification, routing, lightweight agent steps, where token cost dominates the decision. That's Luna, DeepSeek V4-Flash, or Qwen3.8-Max territory, not Opus or Sol. Second: does the task require the deepest available reasoning where a wrong answer is expensive, complex coding, multi-step agentic work, research synthesis. That's where Opus 5, Fable 5, and Sol actually earn their price. Third: do you need to self-host, fine-tune, or audit the weights for compliance reasons. If yes, Kimi K3 and DeepSeek's MIT-licensed release are worth evaluating. (Note added September 8, 2026: Kimi K3's full weights did ship, on July 27, 2026, but under Moonshot's own "Kimi K3 License," an MIT-inspired license with a commercial carve-out requiring a separate agreement for companies above $20M in trailing revenue, not a literal MIT license. Self-hosting also requires substantial infrastructure, on the order of 1.4TB of weights across multiple multi-GPU nodes, so "self-host it yourself" is realistic for well-resourced teams, not solo builders.)

The model wave isn't slowing down, and whatever wins this specific comparison will be old news in another six weeks. The framework above outlasts any single model's benchmark run. Pick based on the job and the volume, not the leaderboard, and you won't have to re-litigate this decision every time a new model drops.