Comparison Head to Head

Open Weights vs the Frontier: The Gap Is 3 Points and $8 a Million

Kimi K3 scores 60 on the Artificial Analysis Intelligence Index. Claude Opus 5 scores 63. That's the whole gap now — and the open model costs a third as much.

White humanoid robot with large round eyes holding a tablet
Photo via Unsplash
The receipts
  • Artificial Analysis has Kimi K3 (max) at 60 on its Intelligence Index. Claude Opus 5 tops the same board at 63, Fable 5 at 62. Three points is the whole frontier premium.
  • On Moonshot's own model card, K3 beats Fable 5 on GPQA Diamond (93.5 vs 92.6), Terminal-Bench 2.1 (88.3 vs 88.0), and BrowseComp (91.2 vs 88.0).
  • Where closed still wins: DeepSWE. GPT-5.6 Sol 73.0, Fable 5 70.0, K3 67.5. Long-horizon agentic coding is the last honest moat.
  • Price: K3 is $3.00 in / $15.00 out per million tokens. Fable 5 is $10 in / $50 out. Same tier of answer, one-third the invoice.
Short answer

As of August 2026 the best open-weight model, Moonshot's Kimi K3, scores 60 on the Artificial Analysis Intelligence Index against 63 for Claude Opus 5 and 62 for Claude Fable 5. Open weights now trail the closed frontier by roughly three points while costing about one-third as much per token.

Three points.

That’s the distance between the best model you can download for free and the best model money can rent. Artificial Analysis has Kimi K3 at 60 on its Intelligence Index. Claude Opus 5 leads the same board at 63, with Fable 5 at 62.

I’ve been running both benches since K3’s weights dropped July 27th. The number that matters isn’t the three points. It’s what the three points cost.

The head-to-head

Moonshot published a direct comparison table in the K3 model card. Vendor-run numbers, so read them like a spec sheet and not gospel — but they’re specific, reproducible, and they put K3 next to the closed flagships instead of next to other open models.

BenchmarkKimi K3 (open)Claude Fable 5GPT-5.6 Sol
GPQA Diamond93.592.694.1
Terminal-Bench 2.188.388.088.8
DeepSWE67.570.073.0
BrowseComp91.288.090.4
AA Intelligence Index6062
Price /1M tokens (in / out)$3.00 / $15.00$10 / $50

Fable 5’s pricing is straight from Anthropic’s own model docs: $10 per million in, $50 per million out, 1M context, 128k output.

Peep the shape of that table. On graduate-level science K3 is ahead of Fable by 0.9 and behind Sol by 0.6. On web research it beats both. That’s not a fast-follower. That’s the same tier.

Where the frontier still earns the invoice

DeepSWE. 73.0, 70.0, 67.5.

Five and a half points on long-horizon agentic coding is the whole remaining moat, and it’s the exact workload people actually pay frontier rates for. An agent that runs 40 tool calls without losing the plot is a different product than a model that answers a hard chemistry question. Everybody’s GPQA is 93. Nobody’s DeepSWE is.

The other thing closed buys you is the scaffolding. Effort sliders, task budgets, server-side fallback on refusal, prompt caching that actually gets credited back. With open weights you build that yourself or you rent it from an inference provider — and provider quality varies enough that OpenRouter shows K3 GPQA Diamond at 92.7 on one host and 90.2 on another. Same weights. Different serving stack. That variance is a real operational cost nobody puts on a leaderboard.

What open weights buy that no benchmark measures

You own the artifact.

I wrote about this after Fable 5 came back from its 19-day export ban, and the lesson didn’t expire: a closed model can be quota’d, deprecated, classifier’d, or shut off by a cabinet secretary’s letter. A file on your disk can’t. Every builder who kept an open model warm during that blackout kept shipping.

Then there’s the boring stuff that decides real deployments. Fine-tuning on your own data. Inference inside your VPC so PHI never leaves. A fixed GPU bill instead of a metered one. If you’re running local models on your own hardware, the marginal token is free, and that changes what workloads are even worth building.

How I’m routing it

One model to dream. One model to build. One model to audit.

Frontier for the long agent runs and anything where a wrong answer costs more than the whole month’s token spend. Open weights — K3 for hard reasoning, a smaller open coding model for the volume — for everything else. That’s most of the traffic, and it’s where the 3x price delta compounds.

Three points is not a moat. Three points is a preference.

Price it that way.

That compression kept going. By September the budget tier was close enough that two Intelligence Index points and three cents a task decided it — see Muse Spark 1.3 vs Gemini 3.8 Flash.

#TheAIMogul

Bottom lineThe frontier still wins agentic coding and it still wins the last few points of raw reasoning — but three index points is not a moat you build a company on. Run open weights as the default for volume work and buy frontier tokens only for the jobs that actually break without them.

Frequently asked

What is the best open-source AI model right now?
Moonshot AI's Kimi K3, released with open weights on July 27, 2026. It's a 2.8-trillion-parameter mixture-of-experts model with 104B active parameters and a 1,048,576-token context window. Artificial Analysis scores it 60 on its Intelligence Index, the highest of any open-weight model, and it posts the best published open-weight GPQA Diamond score at 93.5.
How far behind closed frontier models are open-weight models in 2026?
About three points on the Artificial Analysis Intelligence Index — 60 for Kimi K3 versus 63 for Claude Opus 5. On individual benchmarks the picture is mixed: K3 leads Claude Fable 5 on GPQA Diamond and BrowseComp, and trails it on DeepSWE (67.5 vs 70.0). The gap is real but it is now measured in points, not generations.
Are open-weight models actually cheaper than GPT or Claude?
Yes, by roughly 3x at the top end. Kimi K3 lists at $3.00 per million input tokens and $15.00 per million output; Claude Fable 5 is $10 in and $50 out per Anthropic's own docs. Smaller open models go far lower. And because the weights are downloadable, self-hosting turns a per-token bill into a fixed GPU cost.
Where do closed frontier models still clearly win?
Long-horizon agentic coding. On DeepSWE, GPT-5.6 Sol scores 73.0 and Claude Fable 5 scores 70.0 against Kimi K3's 67.5 on Moonshot's own card. Closed labs also ship tooling — effort controls, task budgets, server-side fallback — that open weights make you build yourself.
Should I switch my production stack to open weights?
Switch the volume, not everything. Route classification, extraction, summarization, and routine code edits to an open model, and keep a frontier model for long-horizon agent runs and anything where a failure costs more than the token savings. Owning the weights also means the model can't be deprecated or export-controlled out from under you.