Writing
The Kumo blog
How token-gateway pricing really works, drop-in migration, and the engineering behind it. No fluff — just the mechanics
Featured
Latest
All articles
View all articles →- GPT-6.1 Terra: OpenAI's new model — capabilities and priceOpenAI released GPT-6.1 Terra, a mid-tier model at half the price of GPT-6.1 Sol with a 1.05M-token context, computer use and big gains on agent tasks. What the measurements show and what it costs on Kumo: $0.9 / $4.3 per million tokens, 15% below OpenAI.gpt-6.1gpt-6.1 terranew modelopenai5 min read
- Claude proved 123 libraries never crash: formal verification with AIAnthropic and Theorem used Claude to formally prove that 239 entry points across 123 open-source libraries never crash on any input under 64 MB. How it works, what the proof actually guarantees, and where contracts and invariants in code come in.anthropicdevelopmentformal verificationresearch10 min read
- Claude Mythos 5 went beyond its test: inside the UK AISI incidentDuring UK AISI cyber testing, Claude Mythos 5 tried to land a malicious pull request in a repository owned by real developers and deceived its maintainers. What happened, why the model did not stop on its own, and which layers of control agents need.ai agentsai safetyanthropicresearch11 min read
- Kimi K3.1 Pro: Moonshot AI's new flagship — capabilities and priceMoonshot AI released Kimi K3.1 Pro, a post-trained open-weight successor to Kimi K3 with a 1M-token context and text, image and video input. What is known about it, how it differs from K3 and what it costs on Kumo: $2.3 / $11.6 per million tokens.kimikimi k3.1 promoonshot ainew model5 min read
- PretrainGym: can an AI agent curate pretraining data on its ownA student at the Prime Intellect RL Residency built an environment where an agent assembles a 400M-token data mix for a small nanoGPT-style model. Kimi K3 beat GLM-5.3 and GPT-6.1 Sol. How the environment works, how cheating was closed off, and the three agents' strategies.ai agentskimi k3researchtraining data10 min read
- The best AI models for coding in 2026: a benchmark rankingWhich model writes code best: a ranking by Terminal-Bench 4.0, Coding Index and Artificial Analysis's coding agents leaderboard as of October 20, 2026. Scores, Kumo prices, the best model for each task and for your editor.benchmarksclaude sonnet 5.5codingrankingvibe coding12 min read
- New AI models you may not have heard of: 43 models from 22 vendorsBesides ChatGPT, Claude, Gemini and Grok, Kumo carries 43 models from 22 vendors: Xiaomi MiMo, NVIDIA Nemotron, StepFun Step, Meituan LongCat, the diffusion model Mercury and more. What makes them interesting, how they score on independent tests, what they cost and how to try them on one key.chinese ai modelsnew modelsnvidia nemotronoverviewxiaomi mimo13 min read
- Refusal directions, abliteration and antislop: what changes inside an LLMA technical explanation of refusal geometry, weight orthogonalization, prompt distillation and GRPO. Read the evidence and explore GLM-5.3 Raw and Qwen3.8 27B Antislop on Kumo.abliterationantislopmodelsresearch11 min read
- Uncensored AI: how uncensored models, abliteration and antislop workWhy models refuse, what abliteration changes, and why antislop targets writing style. Two approaches explained through GLM-5.3 Raw and Qwen3.8 27B Antislop, with current Kumo pricing.abliterationguidemodelsuncensored12 min read
- Claude Code from Russia: install, pay and run without a VPNWhether Claude Code works from Russia, how to install it on Windows, macOS and Linux, which Claude models to connect and what a day of work costs. Pay with a Russian card or SBP from a Kumo balance — no Anthropic subscription, no VPN.claude codeguidepaying from russiavibe coding11 min read
- Gemini 4 Argon review: benchmarks and price of Google's flagshipGoogle's new flagship beats GPT-6 Astra and Claude Opus 5.5 on 12 of 18 vendor tests, yet ranks fifth on the independent index. Google's and Artificial Analysis benchmarks, coding agents, hallucinations and price: $1.1 and $5.5 per million tokens on Kumo.benchmarksgeminigemini 4 argongooglemodel review11 min read
- DeepSeek V4.1 API: prices, benchmarks and how to connectDeepSeek V4.1 Flash is the most used model in programming: one token in four over a week. A 1M context at $0.12 and $0.48 per million tokens on Kumo. Benchmarks, rivals, real-world cost and API setup in five minutes.apibenchmarksdeepseekdeepseek v4.1pricing9 min read
- Grok 4.7: what changed and how to use it on KumoxAI shipped Grok 4.7 — nearly twice Grok 4.6's Terminal-Bench score at Grok 4.6's price of $2 and $6 per million tokens. Kumo pricing, the benchmarks, and two checks that prove the original model is answering.grokgrok 4.7model verificationpricingxai8 min read
- Claude Opus 5.5: what changed and how to use it on KumoAnthropic shipped Claude Opus 5.5 — Fable 5.1's level for $4 and $20 per million tokens. Kumo pricing, the benchmarks, and two checks that prove the original model is answering.anthropicclaudemodel verificationopus 5.5pricing7 min read
- GPT-6 Sol: what changed and how to use it on KumoOpenAI shipped GPT-6 Sol — half the price of GPT-5.6 Sol and close to the best Claude models at a fraction of the cost per task. On Kumo it is $1.1 and $5.5 per million tokens. The benchmarks, and two checks that prove the original model is answering.gpt-6gpt-6 solmodel verificationopenaipricing8 min read
- Cursor in Russia without a VPN: connect Kumo and get back to codeConnect Kumo to Cursor with your own API key: Russian cards, SBP and crypto payments, request checks and a cost comparison with subscriptions.apicursorguidespayments7 min read
- How to connect GPT to Cursor through Kumo after OpenAI's announcementWhat OpenAI's Cursor announcement means, how to connect GPT through Kumo's API, and which custom-key compatibility limits you should check first.apicursorgptguides6 min read
- Low-priced Opus and Fable tokens for developers who code intensivelyOpus 5 and Fable 5.1 prices on Kumo, Anthropic API comparisons and high-volume calculations. No weekly subscription quotas; card or crypto payments.claudefableopuspricing7 min read
- Which models play the best chess — and why it mattersA fast DeepSeek model just beat Claude, GPT and Kimi over the board. Chess Elo looks like a proxy for hard algorithmic reasoning — here’s the debate, and the practical takeaway for picking a model.analysismodels4 min read
- Why an LLM is a risky fact-checker — and how not to get fooledA model has no opinion of its own: it interpolates over its training data and defends its early mistakes. Here’s why frequency beats truth, why a chat gets stubborn mid-conversation, and a practical protocol for validating facts with AI.analysismodels5 min read