DeepSeek & MiMo Are 0.5 Credits Forever — Plus Opus 4.8, Qwen3.7-Max, and 2026's Cheapest Frontier Review Stack

If you have been waiting for “good enough” review to cost less than coffee, this is the release. The mental model we want you to leave with is anchoring in reverse: flagship models still exist for the PR where one missed security bug costs more than a month of credits — but the median PR should not touch them. The median PR should ride open-weight lanes that now score within striking distance of Sonnet-class SWE-bench Verified numbers at 1/20th the credit burn.

Why MiMo landed at 0.5 credits the same week Xiaomi rewrote API pricing

This is not Critique inventing a promo lane in isolation. On May 27, Xiaomi permanently renovated the entire MiMo-V2.5 pricing system: flat per-token rates with no more input-length multipliers, cuts of up to 99% versus the old 256K–1M tiers, Token Plan quotas reset and expanded 5–8×, and a public write-up on how they kept costs down — SWA with SGLang HiCache shrinking KV-cache churn, better expert-parallel bucketing, and higher cache hit rates on long agent runs. Read their announcement in Primary sources below; our 0.5cr / 1cr MiMo floors are the downstream bet that those economics should be the default for PR review, not a weekend experiment.

DeepSeek V4: flagship open weights at volume pricing

DeepSeek’s V4 family was already the rational default for teams that wanted MoE scale and a million-token context class without renting Claude for every specialist pass. What changed is the price story. V4 Flash at 0.5 credits is not “cheap for a demo.” At roughly 1M input + 150k output per credit unit, a deep 5M-token review pass on Flash can land near 2.5 credits — less than a single old-system unit on a mid-tier model. V4 Pro at 1 credit is the open-weight lead for messy PRs: Artificial Analysis quotes GDPval-AA near 1554 Elo on the Pro Max reasoning profile, leading the open-weights pack on agentic work tasks.

Xiaomi MiMo: the other half of the 0.5cr revolution

MiMo v2.5 at 0.5 credits is the parallel bet to Gemma and Ling in our volume tier: Xiaomi positions the Flash lane for high-throughput agent loops, and the tech report cites 73.4% on SWE-bench Verified for the Flash profile while keeping active parameters tiny enough to run wide specialist fan-out. MiMo v2.5 Pro at 1 credit is the escalation lane inside the same family — 78.9% SWE-V in the official card, with Terminal-Bench 2.0 in the high-60s. The vendor-side story matters too: in their May 2026 price-adjustment post, Xiaomi says they “permanently renovate the entire model pricing system,” drop context-length surcharges, and fund the cut with real inference engineering — not a time-boxed coupon. Critique passes that through as permanent catalog pricing so the median PR does not need a flagship model to get a serious second opinion. If your team has been mentally bucketed into “cheap Chinese models = toy reviewers,” update the bucket: the scores crossed the line where selective review becomes irrational.

Frontier additions — Opus 4.8, Qwen3.7-Max, Gemini 3.5 Flash

Claude Opus 4.7 leaves the catalog; Opus 4.8 takes its Ultra slot at 37 credits with a 1M-token context window and Anthropic’s published 88.6% SWE-bench Verified. Opus 4.8 (Fast) uses the same weights at 74 credits — double the floor for teams that buy latency, not capability. Qwen3.6-Max-Preview retires in favor of Qwen3.7-Max at 6 credits: Alibaba’s agent blog cites 80.4% SWE-V and 60.6% SWE-Pro, a cleaner mid-flagship than the old 8cr preview lane. Gemini 3.5 Flash lands at 10 credits as Google’s “near-Pro coding at Flash economics” bet — 55.1% SWE-Pro public, 76.2% Terminal-Bench 2.1 in DeepMind materials.

How to rebuild your policy stack Monday morning

Frequently asked