The snapshot, captured on 30 July 2026 at 09:41 GMT+2, ranks 17 models by real-world token usage on the OpenCode Go gateway. It is not a vendor benchmark or a synthetic test — it is the actual bill developers paid. For European teams trying to plan their August AI budget, two findings stand out: DeepSeek V4 absorbs nearly four out of every five tokens flowing through the gateway, and the cost gap between the cheapest and priciest frontier model has widened to roughly 28× per session.
The numbers behind the headline
By raw token throughput on OpenCode Go, the top of the table belongs to Xiaomi's MiMo v2.5 (961 B tokens), ahead of DeepSeek V4 Flash (628 B), DeepSeek V4 Pro, and a fast-rising GLM-5.2 from Zhipu (628 B tokens, up 38 per cent on the previous period). Moonshot's freshly released Kimi K3 sits at 352 B tokens, with Kimi K2.7-code just below at 296 B.
But raw throughput is not where the story ends. Anomaly's leaderboard cross-references token volume with price per 1M tokens, average cost per session, and a metric that is easy to overlook — cache ratio, the share of input tokens served from cache rather than recomputed. That is where the European reader should look first, because cached input is what turns a "cheap on paper" model into an actually cheap one in production.
Kimi K3 vs GLM-5.2: same fight, different prices
In the "Model Comparisons" section of Anomaly's report, four of the most popular pair comparisons involved Kimi K3 or GLM-5.2. The most-clicked pair is DeepSeek V4 Flash vs DeepSeek V4 Pro — a sibling match — but the second most-clicked is MiMo v2.5 vs GLM-5.2, and the third is DeepSeek V4 Pro vs MiMo v2.5. GLM-5.2 also appears paired against Kimi K3 in adjacent-leaderboard comparisons. The pattern is clear: developers are shopping for a premium-context model and weighing two Chinese up-and-comers against the DeepSeek default.
Here is what the shopping list looks like, with EUR figures converted today at 1 USD ≈ 0.92 EUR:
- DeepSeek V4 Flash — $0.28 / €0.26 per 1M input tokens, $0.089 / €0.082 average cost per session, 95 per cent cache ratio.
- Xiaomi MiMo v2.5 — $0.28 / €0.26 per 1M input, $0.086 / €0.079 per session, 92 per cent cached.
- MiMo v2.5 Pro — $0.87 / €0.80 per 1M input, $0.234 / €0.215 per session.
- MiniMax M3 — $1.20 / €1.10 per 1M input, $0.720 / €0.662 per session, 93 per cent cached.
- Qwen 3.7 Plus — $1.60 / €1.47 per 1M input, $0.543 / €0.499 per session.
- Kimi K2.6 — $4.00 / €3.68 per 1M input, $0.878 / €0.808 per session, 92 per cent cached.
- Kimi K2.7-code — $4.00 / €3.68 per 1M input, $1.124 / €1.034 per session.
- GLM-5.2 — $4.48 / €4.12 per 1M input, $2.472 / €2.274 per session — the highest on the leaderboard — 76 per cent cached (the lowest among the top 11).
- Kimi K3 — $15.00 / €13.80 per 1M output tokens, $1.876 / €1.726 per session, 93 per cent cached.
Two things jump out. First, GLM-5.2 is about 16× more expensive per million input tokens than DeepSeek V4 Flash — and roughly 27× more expensive on a per-session basis, because its cache ratio of 76 per cent means the average request re-reads more fresh context than competitors. Kimi K3, despite a notable $15.00 per 1M output price, actually ends cheaper per session ($1.88) because the gateway reports a stronger 93 per cent cache hit on it.
Second: when token throughput and price are combined, DeepSeek V4 Flash is so cheap that even though it processed 628 B tokens — less than Xiaomi MiMo v2.5's 961 B — it still accounts for the bulk of DeepSeek's 77.4 per cent vendor token share on the platform. The Anomaly chart also shows a +596 per cent surge in Tencent's slice, hinting Hunyuan-backed offerings may be entering the gateway's model mix off the visible top-17 list.
Why the cache ratio column matters for European teams
For a CTO in Berlin or a consultancy in Madrid planning 10,000 daily agentic sessions, a "headline" price of $0.28 per 1M tokens hides nothing, but a hidden 76 per cent cache ratio means almost a quarter of the prompt is reprocessed each turn. Over a working day that gap compounds: 1000 sessions of GLM-5.2 cost €2,274 versus just €82 for the same volume on DeepSeek V4 Flash — a 27× difference, all from numbers captured in the leaderboard Anomaly published this morning.
This is also why we keep returning to OpenCode Go-analogue data in our own AI Arena benchmarks: per-token prices tell you nothing about real cost until you measure the cache ratio. Kimi K3 looks expensive on a per-output-million basis, but in practice it ends cheaper per session than GLM-5.2 because the gateway successfully reuses 93 per cent of its prompt cache — a finding consistent with what we measured when Kimi K3 first went open-weight earlier this month.
The European angle: who actually uses these models here?
The Anomaly snapshot also breaks down token usage by country. The leaders are China (19 per cent) and the United States (15 per cent), but Europe is not a rounding error: Germany 5.8 per cent, Spain 4.4 per cent, France 3.4 per cent, United Kingdom 2.6 per cent, Italy not far behind. Add the EU-27 plus UK and the European share approaches one quarter of all traffic on OpenCode Go.
That is meaningful for two reasons. First, every one of those European sessions flows through an inference gateway operated outside the EU. Under the EU AI Act — which entered its omnibus enforcement package six days ago with its December deadline for high-risk providers — companies that embed a third-party foundation model in a product sold in the bloc must be able to report which model they used, what data it saw, and where inference happened. A gateway like OpenCode Go, which routes prompts to Chinese model providers, is exactly the kind of supply chain European procurers need to document.
Second, under GDPR Article 28 (processor obligations), sending European users' prompts to a Chinese inference provider raises legitimate transfer-impact-assessment questions. None of the Chinese vendors on the Anomaly leaderboard — DeepSeek, MiniMax, Moonshot, Zhipu, Xiaomi, Tencent, Qwen — currently publishes a verifiable EU adequacy decision or a Standard Contractual Clauses wrapper for their API offering. European firms integrating them should be treating those calls as they would any other non-adequate third-country data transfer.
What this means in practice
For developers and small European teams, three practical takeaways emerge from the Anomaly leaderboard:
- Default to DeepSeek V4 Flash for high-volume, latency-tolerant tasks — the 95 per cent cache ratio and €0.26 per 1M input make it a rational budget option. Use Kimi K3 or GLM-5.2 only when you needle bigger context, multi-step reasoning, or HTML/structured output.
- Track cache ratio, not headline token price — GLM-5.2's 76 per cent cache ratio turns a $4.48 sticker price into a $2.47 per-session cost. That invisible multiplier is where European budgets go missing.
- Document your model supply chain for the AI Act — if your product lands in EU-facing use this autumn, you should be able to show an auditor which model handled each request, where it was inferred, and under what data-transfer arrangement. Gateways like OpenCode Go can provide request logs, but the contractual wrapper is yours to assemble.
This whole story is, in a sense, a sequel to two we have already written: our deep-dive on GLM-5.2 when Zhipu opened its weights, and our coverage of Moonshot's Kimi K3 launch and infrastructure. Both articles predicted that real-world spending would migrate toward whichever vendor could make cache work hardest. Today's leaderboard is the first clean confirmation that, on OpenCode Go at least, that prediction held: DeepSeek won the volume race, and Kimi K3 surprised by winning the per-session efficiency race against its more expensive Chinese rival.
The August model wave — GLM-5.5, Qwen 3.8 Max, DeepSeek V4 refinements — will land within weeks. The leaderboard already shows Zhipu's up 38 per cent and Tencent up nearly 600 per cent week-on-week. The pricing equilibrium on this gateway is going to move again before summer ends.
Is OpenCode Go itself available in the EU?
Yes — OpenCode Go is a public inference gateway reachable from European networks, and Anomaly's geo breakdown shows Germany, Spain, France, and the UK already contributing a combined ~15 per cent of its token traffic. EU firms should treat it as any non-adequate third-country data transfer: log the model used, route sensitive prompts carefully, and keep a transfer-impact assessment on file.
If Kimi K3 reports $15.00 per 1M output tokens, why is its per-session cost only $1.88?
Because the OpenCode Go gateway reports a 93 per cent cache hit on Kimi K3, meaning most input tokens are served from cache rather than reprocessed. Output still costs $15.00 per 1M, but the average session produces only a small fraction of that million. In large part, the model's sticker price looks high on a per-token basis but is competitive on the per-session basis developers actually pay.
Where can I see the Anomaly Innovations leaderboard myself?
Anomaly Innovations publishes the OpenCode Go leaderboard publicly on its model-data portal, with daily snapshots at 09:41 GMT+2. The figures quoted here correspond to the snapshot dated 30 July 2026.
