1. Chinese open-weight models now process the majority of tokens on OpenRouter — north of 20 trillion a week. This stopped being a hobbyist story.
2. The frontier still belongs to Anthropic. Claude Fable 5 leads SWE-bench Pro at 80.3%, and nothing Chinese is within 18 points of it.
3. Alibaba closed its frontier tier. "Qwen = open weights" is no longer true, and Meta has effectively left the race entirely.
4. Self-hosting to save money is almost always a mistake at your volume. The maths is brutal and it's in this chapter.
One warning before the numbers. SWE-bench Verified and SWE-bench Pro are not the same benchmark and are not comparable. Pro is far harder — roughly, 80% on Verified corresponds to about 60% on Pro. Labs have started quoting whichever one flatters them, so any table that mixes the two without saying so is lying to you. This one says so.
Start here, because you use it daily and because the lineup genuinely confuses people. There is no "Opus 5." A new tier name — Fable — now sits above Opus, while Sonnet has jumped a whole generation number and Opus is still on 4.x. Ranked by capability:
| Tier | Model | What it's for |
|---|---|---|
| Frontier | Claude Fable 5 | The most capable widely-released model in the world. Long-horizon agents, hardest reasoning |
| Frontier (gated) | Claude Mythos 5 | Fable 5 without safety classifiers. Invitation-only, defensive cyber. Not self-serve |
| Workhorse | Claude Opus 4.8 | Anthropic's own recommended default for complex agentic coding. This is what you're running now. |
| Default | Claude Sonnet 5 | Best speed/intelligence blend. Default on Free and Pro |
| Cheap & fast | Claude Haiku 4.5 | Low-latency, cost-sensitive. Noticeably older — Feb 2025 cutoff, 200K context |
| Fable 5 | Opus 4.8 | Sonnet 5 | Haiku 4.5 | |
|---|---|---|---|---|
| Context | 1M | 1M | 1M | 200K |
| Price in / out | $10 / $50 | $5 / $25 | $3 / $15 intro $2/$10 to 31 Aug | $1 / $5 |
| Knowledge cutoff | Jan 2026 | Jan 2026 | Jan 2026 | Feb 2025 |
| SWE-bench Pro | 80.3% ✱ | 69.2% | 63.2% | — |
✱ Best score of any model tested, open or closed.
Opus 4.7 introduced a new tokenizer, now used by Fable 5. The same text produces roughly 30% more tokens than on pre-4.7 models. Your 1M-token window is not as big as the number suggests, and your bill is not comparable to last year's. Nobody advertises this.
Refusals return HTTP 200, not an error. Fable 5 can come back with stop_reason: "refusal" — a successful response containing a refusal. If your code only handles error status codes, you will silently treat a refusal as a valid completion. You are not billed for a pre-output refusal, and there is a server-side fallback parameter, but you have to actually wire it up.
Fable 5 sometimes silently routes to Opus 4.8. When a query touches cyber, bio, chem, or model-distillation topics, you may get the weaker model without being told. Under 5% of sessions — but you paid Fable prices.
Worth knowing because it explains behaviour you may have noticed. On 12 June 2026, US export controls hit Fable and Mythos after a research report demonstrated bypassing safeguards to find software vulnerabilities. Anthropic suspended access for roughly nineteen days. Controls lifted on 30 June; the model returned globally on 1 July with a new cyber classifier — and Anthropic itself acknowledged that the new classifier flags benign coding and debugging requests more often than before. If Claude has felt twitchier about security-adjacent code in the last two weeks, that is why, and it is not your imagination.
The Gemini line is in an odd state: the newest Flash is a generation ahead of the newest Pro. There is no Gemini 3.5 Pro.
| Model | In / Out per 1M | Note |
|---|---|---|
| Gemini 3.5 Flash (May 2026) | $1.50 / $9.00 | Beats 3.1 Pro on most coding and agentic benchmarks, ~4× faster |
| Gemini 3.1 Pro (Feb 2026) | $2 / $12 (≤200K) $4 / $18 (>200K) | 80.6% SWE-bench Verified · 94.3 GPQA-D · 1M context |
| Gemini 3 Flash | $0.50 / $3.00 | Cheapest credible Google tier |
| Gemini Omni Flash | $1.50 / $9.00 $17.50 video out | See below — it is not what you'd assume |
For agent work, reach for the Flash, not the Pro. That inverts the usual intuition and it is the single most useful thing to know about Google's lineup right now.
It got a lot of headlines at I/O 2026 and it is real — but it is a generative world model for any-input-to-any-output media, starting with video, edited conversationally instead of on a timeline. Demis Hassabis framed it as a world model with physics understanding, not a video generator. It ships in the Gemini app, Google Flow, and YouTube Shorts Remix.
Do not plan a coding-agent migration around it. It is not competing with Sol or Opus.
DeepSeek V4 (March 2026) ships two models, both 1M context, both MIT-licensed open weights:
| Model | Input | Output | Note |
|---|---|---|---|
deepseek-v4-pro | $0.435 | $0.87 | 1.6T params MoE · 80.6% SWE-bench Verified — ties Gemini 3.1 Pro |
deepseek-v4-flash | $0.14 | $0.28 | The price floor of the entire industry |
Sit with those numbers for a second. DeepSeek V4-Pro's output costs $0.87 per million tokens. Claude Fable 5's costs $50.00. That is a 57× ratio, for a model roughly fifteen SWE-bench points behind. V4-Flash against GPT-5.6 Sol is 107×.
deepseek-chat and deepseek-reasoner are deprecated on 24 July. They map to the non-thinking and thinking modes of deepseek-v4-flash. If either string appears anywhere in your code, that is a this-week problem.
(The long-awaited R2 reasoning model never shipped — reporting says the CEO was unsatisfied with it. The reasoning capability was folded into V4's thinking mode instead.)
Kimi K2.6 (April 2026), 1T params MoE, effectively MIT-licensed, $0.95 / $4.00, 256K context, 80.2% SWE-bench Verified.
Kimi ships an architecture that runs up to 300 sub-agents across ~4,000 coordinated steps for 12+ hours of continuous autonomous execution. Nothing else — open or closed — ships an equivalent.
If you ever seriously pursue the "let it run overnight on the repo" pattern you sketched for Cinefolio, this is the only model in the world explicitly built for it. Worth a look on that project specifically.
Z.ai's GLM-5.2 (June 2026), ~750B MoE, MIT, 1M context, $1.40 / $4.40. It scores 62.1 on SWE-bench Pro — beating GPT-5.5's 58.6 — and is the top open model on the Artificial Analysis Intelligence Index. It landed within one percentage point of Claude Opus 4.8 on a closely-watched agentic benchmark at roughly a fifth of the cost. There is an $18/month coding plan.
Alibaba closed its frontier tier. Qwen 3.5 and 3.6 remain Apache 2.0 — but Qwen3.7 Max is closed, API-only. The biggest Chinese open-weights champion took its best model private. Stop assuming "Qwen" means "open."
Meta has effectively left. Llama 4 Behemoth was previewed, delayed, and shelved — never released. In April 2026 Meta Superintelligence Labs shipped Muse Spark, which is closed-weight and consumer-facing (meta.ai, WhatsApp, Instagram, the glasses). There is no serious first-party developer API story. The company that started the open-weights movement is no longer in it.
Nemotron 3 Ultra — 550B hybrid Mamba-2/Transformer MoE, OpenMDW license, $0.42 / $2.61, and the strongest US open-weight model in existence. NVIDIA also launched a "Nemotron Coalition" of global labs for open frontier models. If you want frontier-adjacent open weights that are not from China — for a client with procurement rules, say, which given your ERP work is a real scenario — this is now the answer, and it got almost no press.
USD per million tokens, standard tier. SWE-bench Verified unless marked Pro — again, they are not comparable.
| Model | Company | Open? | Ctx | In | Out | SWE-bench | Standout strength |
|---|---|---|---|---|---|---|---|
| Claude Fable 5 | Anthropic | Closed | 1M | 10.00 | 50.00 | 80.3% Pro ✱ | Best coding model, period |
| Claude Opus 4.8 | Anthropic | Closed | 1M | 5.00 | 25.00 | 69.2% Pro | Anthropic's agentic default — yours |
| Claude Sonnet 5 | Anthropic | Closed | 1M | 3.00 | 15.00 | 63.2% Pro | Best price/perf in the Claude line |
| Claude Haiku 4.5 | Anthropic | Closed | 200K | 1.00 | 5.00 | — | Fastest — but stale (Feb 2025 cutoff) |
| GPT-5.6 Sol | OpenAI | Closed | 1.05M | 5.00 | 30.00 | 64.6% Pro | ARC-AGI-3 SOTA; terminal/browser agents |
| GPT-5.6 Terra | OpenAI | Closed | 1.05M | 2.50 | 15.00 | n/p | Best all-round default; ~Sol at ½ price |
| GPT-5.6 Luna | OpenAI | Closed | 1.05M | 1.00 | 6.00 | n/p | Cheap. ⚠️ 41.3% long-context recall |
| Gemini 3.5 Flash | Closed | 1M | 1.50 | 9.00 | n/p | Beats its own Pro at coding, 4× faster | |
| Gemini 3.1 Pro | Closed | 1M | 2.00 | 12.00 | 80.6% | Long-context; 94.3 GPQA-D | |
| DeepSeek V4-Pro | DeepSeek | MIT | 1M | 0.435 | 0.87 | 80.6% | Frontier-adjacent coding at 1/57th Fable's output cost |
| DeepSeek V4-Flash | DeepSeek | MIT | 1M | 0.14 | 0.28 | n/p | The industry price floor |
| Kimi K2.6 | Moonshot | ~MIT | 256K | 0.95 | 4.00 | 80.2% | Agent Swarm — 300 sub-agents, 12hr runs |
| GLM-5.2 | Z.ai | MIT | 1M | 1.40 | 4.40 | 62.1% Pro | Top open model on AA index; beats GPT-5.5 |
| Qwen3.7 Max | Alibaba | Closed | 1M | 2.50 | 7.50 | n/p | Multi-hour autonomous agent; GPQA 92.3 |
| Qwen 3.5 | Alibaba | Apache 2.0 | — | self-host | n/p | Best truly open Qwen; 91.3 AIME | |
| Grok 4.5 | xAI | Closed | 500K | 2.00 | 6.00 | n/p | #1 agentic tool use; 2× token efficiency. No EU |
| Mistral Large 3 | Mistral | Apache 2.0 | 262K | self-host | n/p | Best multilingual; EU-sovereign | |
| Nemotron 3 Ultra | NVIDIA | OpenMDW | — | 0.42 | 2.61 | n/p | Strongest US open-weight model |
| Muse Spark | Meta | Closed | — | consumer only | n/p | No real developer API. Meta is out | |
✱ Best of any model tested. n/p = not published by the vendor — if someone quotes you one, they're guessing.
This question comes up constantly and the honest answer is unpopular, so here is the arithmetic.
| Break-even against… | You'd need to be running… |
|---|---|
| A premium API (~$5/Mtok) | ~256 million tokens/month, at 60–70% sustained GPU utilisation |
| A cheap open API like DeepSeek V4-Flash | ~5.7 billion tokens/month on a single H100 |
Read the second row again. Self-hosting DeepSeek to beat DeepSeek's own API price is essentially never worth it. And raw GPU cost is only 30–40% of true total cost of ownership — budget a 2.5–3× multiplier, because engineer time is another 25–30%, and a senior inference engineer runs $250–360K loaded.
Self-host for data residency, compliance, or fine-tuning — never to save money. At the volumes any of your projects run, the open-weight APIs (DeepSeek, Moonshot, Z.ai, Together, Fireworks) capture essentially 100% of the cost advantage with 0% of the operations burden.
The one place this could genuinely matter to you: an Iraqi client with data-residency requirements. That is a compliance argument, not a cost one — and it's the right reason to reach for it.
The strategic move in 2026 is not picking a model. It is routing by task, because the cost spread is now 50× while the quality spread on easy tasks is nearly zero.
| When you're doing… | Reach for… | Why |
|---|---|---|
| The gnarly refactor you can't afford to get wrong | Fable 5 / Opus 4.8 | 80.3% SWE-bench Pro is not matched anywhere, at any price |
| Everyday agentic coding | Opus 4.8 or GPT-5.6 Terra | Both land near the frontier at a fraction of flagship cost |
| Bulk codegen, test generation, classification, anything in a loop | DeepSeek V4-Flash, GLM-5.2, Kimi K2.6 | The 20–60× cost delta dwarfs the quality delta on easy tasks |
| Overnight autonomous runs on a big repo | Kimi K2.6 | Agent Swarm — nothing else does 12-hour continuous execution |
| A client who cannot use Chinese weights | NVIDIA Nemotron 3 Ultra | Strongest US open-weight option |
The top of the frontier is still American. Nothing Chinese matches Fable 5's 80.3% on SWE-bench Pro — the best is GLM-5.2 at 62.1, an eighteen-point gap on the hardest agentic coding benchmark. That gap is worth $50 per million tokens to some teams and precisely nothing to others. Know which one you are before you switch.
Sources. Anthropic platform docs & Fable 5 announcement · OpenAI developer docs · Google Gemini API pricing & I/O 2026 announcements · DeepSeek API docs & HuggingFace · Moonshot Kimi platform docs · Z.ai / VentureBeat · Artificial Analysis · Vellum · OpenRouter open-weights report (June 2026) · CNBC on Chinese model adoption · NVIDIA Nemotron Coalition · Mistral · x.ai. Prices and scores current as of 14 July 2026 and moving fast; the live version of this chapter is kept updated.