Open weights vs. closed models: should you switch?

TL;DR

1The two camps

Start with the word, because it does real work here. Open-weight means the parameters are published and you can download, host and fine-tune them. It does not mean open source: the training data and pipeline stay private in every case below, and the licenses range from plain MIT to bespoke agreements with revenue gates. Closed means the model exists only behind an API you rent. The practical difference is not ideology. It is whether the model can be taken away from you, whether you can run it on your own hardware, and who captures the margin on every token.

The closed side

Anthropic ships Haiku 4.5 for cheap fast tasks, Sonnet 5 as the volume workhorse, Opus 5 for hard agentic coding, and Fable 5 at the top. Opus 5 leads Artificial Analysis's Intelligence Index at 63 and its Agentic Index at 55.3, and holds the coding crown on both of Arena's vote-based boards. OpenAI's GPT-5.6 Sol narrowly leads the overall composite snapshot and is the better generalist. Google's Gemini 3.1 Pro leads the reasoning benchmarks. xAI's Grok 4.5 takes the near-frontier value slot, at 91% of the leading score for 88% less output cost. The separation across all four is a few points on most composites, which is why the real question is rarely which is smartest.

What the closed labs actually sell on top of the weights is the harness: adaptive thinking with effort control, prompt caching, server-side tools, compaction, batch pricing, and a first-party API the whole toolchain is tuned against. That is the part the open camp cannot copy by publishing a checkpoint.

The open side

There is no single open-weight answer any more, which is the main thing that changed this year. Five models matter, and they are good at different things:

• Kimi K3 — Moonshot's flagship, 2.8T parameters with 104B active, weights published 27 July. Natively multimodal, 1M context, best open model overall on agentic coding at 88.3 on Terminal-Bench 2.1 and 81.2 on FrontierSWE. The catch is the license, which is a bespoke Kimi K3 agreement rather than Apache or MIT, and the footprint: roughly 1.6 TB of weights.
• DeepSeek V4 — plain MIT, no revenue gate, no attribution clause. Ships as V4-Pro (1.6T total, 49B active) and V4-Flash (284B total, 13B active). Text-only, and it will lose the hardest runs, but it sets the price floor everything else is measured against.
• GLM-5.2 — Zhipu's long-horizon coding model, MIT-licensed, 1M context, 62.1% on SWE-bench Pro. The strongest model available under a genuinely permissive license, and the one to reach for when the license itself is the constraint.
• MiniMax M3 — the first open-weight model to combine coding, native multimodal input and computer use at 1M context in one checkpoint.
• Qwen3.6 — Alibaba's range, spanning sizes that run on a laptop up to server tiers, under permissive licenses. The practical choice for local and on-premises work.

Two more names show up at the top of August's open-weight boards and are worth knowing even though they sit outside this comparison: NVIDIA's Nemotron 3 Nano Omni, and the growing tier of community fine-tunes and quantizations that appear within days of any major weight drop.

2How big is the gap

The useful way to size this is not a benchmark table but a clock. Epoch AI tracks both camps on its Capabilities Index and reports the best open-weight models running about four months behind the closed frontier, an 8-point ECI gap — roughly the distance between consecutive point releases of the same model family. When Epoch last cut the numbers, GLM-5 sat at 146.62 and Kimi K2.6 at 151.60, against GPT-5.5 Pro at 159.35 and Claude Opus 4.7 at 156.18.

Four months is the whole argument. It is short enough that most work does not notice, and long enough that the hardest work does. The pattern across independent evaluations is consistent: open weights close 70–90% of the capability gap at 5–10× lower cost per token, and the part they do not close is specific rather than general. It shows up in long-horizon agentic reasoning — holding coherence over a long chain, noticing a wrong turn, backtracking cleanly. That is exactly the profile of repo-scale debugging and multi-hour agent runs, and exactly why the verdict at the end is a routing rule rather than a winner.

On the individual boards the picture is genuinely mixed. Kimi K3 placed top-three across six coding benchmarks at launch and took #1 on Arena.AI's Frontend Code Arena, ahead of both Claude Fable 5 and GPT-5.6 Sol. Claude Opus 5 still leads Artificial Analysis on raw intelligence and agentic work, and generates substantially faster.

Inside the open camp: K3 against DeepSeek V4

Choosing between open models turns out to be harder than choosing between the camps, because the spread inside the open camp is wider than the gap to the frontier. DeepSeek shipped V4 on April 24, 2026 in two tiers — V4-Pro (1.6T total, 49B active) and V4-Flash (284B total, 13B active) — both MIT-licensed, both with 1M-token context. Set against K3 the comparison splits along a line that matters more than any headline score: writing code versus doing coding work.

V4-Pro is the best model in the world at single-shot algorithmic coding — LiveCodeBench 93.5, Codeforces Elo 3,206 — and was never a threat at agentic work: 12.8 on DeepSWE against K3's 67.5. A model that wins competitive programming and loses at navigating a repo over forty tool calls is telling you those are different skills. For Claude Code-shaped work, V4-Pro was not the competition.

Then V4-Flash-0731 landed on July 31 — one day after this piece's cutoff — re-post-trained specifically for agents. It beats the V4-Pro preview on all nine benchmarks DeepSeek publishes, by 8.7 to 41.6 points: Terminal-Bench 61.8 → 82.7, DeepSWE 7.3 → 54.4. The small tier now out-agents the flagship, and it sits 5.6 Terminal-Bench points behind K3 at 2% of K3's output price. That, not V4-Pro, is the comparison worth running.

The long view: capability over time

Zoom out and the question looks different. On SWE-bench Verified — the share of real GitHub issues a model can resolve end-to-end — frontier models went from a third of issues in mid-2024 to over 90% today. The line to watch is the aqua one: open-weight models trailed the frontier by 20–30 points for two years. DeepSeek V4-Pro narrowed that to about seven points in April 2026, and K3 closed it almost entirely three months later.

3What it costs

Note what the two blocks have in common: every model here runs a 1M-token context window except Haiku. Context length stopped being a differentiator this year, which is part of why the argument moved to price and licensing.

Artificial Analysis puts the blended real-world cost at $2.31 per 1M tokens for K3 versus $3.85 for Opus 5 at high effort — roughly a 40% saving against the model most people are thinking of leaving. Fable 5 doubles Opus again. The question isn't whether K3 is cheaper per token (it is); it's how many Opus-or-Fable-grade tasks you'd feed to a model that fails them. Price the retries, not the tokens.

K3 is not the price floor, though, and the table above is the first place that shows. Take a day of agentic coding at roughly 2M input and 200K output tokens, before caching: Opus 5 runs $15.00, K3 runs $8.80, V4-Pro runs $1.04, and V4-Flash runs $0.34. Agentic work is output-heavy in the expensive direction, and output is where the 17× and 54× gaps live. The honest framing is that K3 buys you roughly 40% off Opus while staying in the frontier pack; DeepSeek buys you an order of magnitude off both, and you pay for it in capability — a trade the next section prices out.

The long view: price over time

Capability climbed while price collapsed. GPT-4 launched at $30 per million input tokens in March 2023; today's flagships sit at $5, and yesterday's frontier is nearly free — Epoch AI estimates GPT-4-level performance has fallen from ~$20 to ~$0.40 per million tokens. Note where K3 sits: at $2.90 it is expensive for an open-weight model — Kimi's own K2 launched at a fifth of that, and DeepSeek's V4-Pro, three months older than K3, lists at $0.435 — because K3 is priced as what it is, a frontier model that happens to publish its weights. The aqua line bending back up in 2026 is that decision, made visible.

4Where each camp wins

5The switching path

The reason this question is even live: Claude Code doesn't care whose model answers it. It reads a base URL, a token, and a model name from the environment, and Moonshot ships an Anthropic-compatible endpoint built specifically to be a drop-in target. Switching is three variables, not a new tool:

# Point Claude Code at Kimi K3 — Moonshot first-party or OpenRouter,
# one API key from either console is enough.
export ANTHROPIC_BASE_URL="<anthropic-compatible endpoint>"
export ANTHROPIC_AUTH_TOKEN="<your Moonshot or OpenRouter key>"
export ANTHROPIC_MODEL="kimi-k3"
claude   # same harness, different brain

Get the endpoint URL and a key from Moonshot's console (platform.kimi.ai) or from OpenRouter. Unset the three variables — or keep them in a separate shell profile — and you're back on Claude. Switching cost is near zero in both directions.

On OpenRouter, pick the routing mode deliberately: Exacto routes for tool-calling accuracy, which is what an agentic harness needs; Nitro optimizes speed, Balanced price. Flaky tool calls that look like model weaknesses are often just provider variance.

6What you give up

• The harness is tuned for Claude. Claude Code's prompts, tool definitions, and thinking integration are built against Anthropic models. K3 through a compatibility shim runs well, but adaptive thinking, effort levels, server-side tools, compaction, and prompt-cache economics either change behavior or disappear. Expect to lose polish even where raw capability is equal.
• Throughput. At ~33 tokens/sec vs ~54, long generations feel meaningfully slower, even though K3 starts answering sooner.
• A single accountable vendor. First-party Anthropic API vs. a routing layer over multiple hosts. When something breaks at 2am, "which provider served that request" is a real question on OpenRouter and a non-question on Anthropic.
• Ecosystem gravity. Managed agents, fallbacks, batch pricing, the models API — the supporting features accumulate on the first-party side.

And what you gain beyond price: weights you control. No deprecation schedule, no retention policy you didn't write, fine-tuning on your own data, and the option — if you ever have the hardware — of running the thing yourself.

7What happens next

Everything above is a snapshot. K3's weights are four weeks old, MiniMax M3 is newer than that, and Anthropic's Sonnet 5 intro pricing ends August 31. Choosing a camp today is a bet on where both curves go. Six forecasts, each with the signal that confirms or kills it, reviewed against what the last month actually did.

Forecasts made 30 July 2026, reviewed 24 August; confidence labels are judgment, not measurement. The macro side of this — what commodity inference does to an AI buildout financed on off-balance-sheet debt — is tracked weekly in the newsletter.

8The verdict

Nobody sensible picks a camp. They split the work across both, and the split is legible once you stop asking which model is better and start asking what a failure costs.

• Open weights for the routine tier — high-volume, low-stakes coding, where the four-month capability gap does not show and a 5–10× discount compounds every day.
• Open weights for visual work — UIs, dashboards, games, anything where the model iterates against screenshots. K3 holds #1 on the Frontend Arena and MiniMax M3 brings computer use, so this is a camp strength now rather than one model's trick.
• Open weights when the requirement is structural — self-hosting, data residency, air-gapped deployment, or a model nobody can deprecate. The closed labs have no answer to this, and it is the one dimension where the gap is infinite rather than four months.
• Closed for long-horizon work — repo-scale debugging and multi-hour agent runs. This is precisely where the residual gap lives: holding coherence over a long chain and backtracking cleanly. Opus 5 and Fable 5 remain the stronger defaults.
• Read the license before the benchmark — “open weights” spans plain MIT (DeepSeek V4, GLM-5.2) and bespoke agreements with revenue gates (Kimi K3). If you plan to re-host or fine-tune, that clause decides more than any leaderboard position.
• Go to the floor if cost binds — DeepSeek V4-Flash-0731 does agentic coding within six Terminal-Bench points of K3 at 2% of its output price, MIT, no revenue gate. Text-only, and it will lose the hard runs, which is exactly why it belongs on the routine tier rather than the whole job.
• Either way, run the experiment: point the harness at one open model for a week of real tasks and keep your own pass/fail notes. Switching costs three environment variables in each direction. Not testing is the only wrong answer.

Sources

• VentureBeat — Moonshot AI releases Kimi K3, the largest open-source model ever
• Tom's Hardware — 2.8-trillion-parameter Kimi K3 beats Claude Fable 5 in Frontend Code Arena
• OpenRouter — Kimi K3 pricing & specs
• Artificial Analysis — Claude Opus 5 (high) vs Kimi K3
• NxCode — Kimi K3 benchmarks explained
• Kimi API Platform — Use Kimi in Claude Code
• LLM Gateway — Kimi K3 with Claude Code, Cursor, and Cline
• CodingFleet — Claude Opus 5 vs Kimi K3 benchmark comparison
• Epoch AI — LLM inference prices have fallen rapidly but unequally
• TokenMix — AI API pricing history, GPT-4 to 2026
• CodeAnt — SWE-bench leaderboard 2026
• MorphLLM — Claude benchmark scores & SWE-bench Verified history
• DeepSeek — models & API pricing (V4-Pro, V4-Flash)
• Hugging Face — what changed in the DeepSeek V4-Flash-0731 build
• Caixin — DeepSeek releases official V4-Flash model
• MarkTechPost — Kimi K3 vs DeepSeek V4-Pro vs GLM-5.2 on benchmarks, license, and serving cost
• DeepInfra — Kimi K3 vs DeepSeek V4-Pro vs GLM-5.2
• Artificial Analysis — Kimi K3 vs DeepSeek V4-Pro

Claude pricing and model specs from Anthropic's published API pricing as of July 2026. DeepSeek figures added August 5, 2026.