DeepSeek just crashed the code price — and the AI coding stack reshuffled this week

DeepSeek crashed the price floor. On May 22-23, 2026, DeepSeek made its 75% V4-Pro discount permanent. V4-Flash sits at $0.14 per million input tokens and $0.28 per million output — roughly 35 times cheaper than GPT-5.5 and approximately 100 times cheaper than Claude Opus 4.7 on output. The Decoder confirmed: DeepSeek's output tokens are now priced at least 34x below GPT-5.5. And cache hits got hammered too — input cache price dropped to 1/10 of launch, meaning V4-Flash cached input is $0.0028 per million tokens. A 98% reduction. This is not a promotional moment. This is the new pricing structure of AI coding.

The mid-tier just absorbed the frontier. Cursor's Composer 2.5 launched May 18 on Moonshot's Kimi K2.5 base. SWE-Bench Multilingual 79.8% — on par with Claude Opus 4.7 and GPT-5.5. Priced at $0.50 per million input tokens. Roughly 1/10 the per-task cost of the frontier alternatives. Qwen 3.7 Max landed May 20-21 at Intelligence Index 56.6, ranking 5th globally and first in China, ahead of Gemini 3.5 Flash. Priced at $2.50 per million input. The new mid-tier is doing frontier-grade coding for sub-frontier dollars.

Anthropic's roadmap leaked. The Information's exclusive on April 14 — sourced from a March 31 second-wave CMS breach — exposed Claude Opus 4.7 (since released April 16), unreleased Sonnet 4.8, and a mysterious model codenamed "Capybara." Anthropic has never historically shipped Sonnet two minor versions ahead of Opus, so Sonnet 4.8 is the next imminent ship. Capybara is the wildcard — internal docs reportedly tied it to autonomous workflow tooling beyond what's currently shipping in Claude Code. As of May 24, Sonnet 4.8 has not appeared in the API or in the Claude.ai model selector.

OpenAI's next move is half-leaked too. GPT-5.6 canary signals went live in mid-May 2026. ChatForest tracked the canary endpoints; Polymarket priced May release at low probability but June-July at a coin flip. Expected feature set: 1.5M context window, "Codex UltraFast" mode, sharpened agentic planning, and material improvement in error recovery — the ability for the model to recognize a failed step and self-correct without human intervention. None of these specs are confirmed. But OpenAI's Codex with GPT-5.5 already hit Terminal-Bench 2.0 at 82.7% and SWE-Bench Pro at 58.6% — the current public state-of-the-art for autonomous coding.

The composable coding stack is the new default. Three of the most-shipping engineering teams in May 2026 are running Claude Code + Codex CLI + DeepSeek TUI in three separate terminals — behind one API key — and switching by task class. Claude for architectural changes and complex multi-file refactors. Codex for autonomous batch tasks and long-horizon agent work. DeepSeek for everything where token cost matters more than peak quality. Cursor sits in the IDE for daily feature work. Composer 2.5 has folded that into a single product that can match the frontier at a fraction of the cost.

The strategic read. This week confirmed that the AI coding war is no longer about which model is the best. It is about price-per-task at a given quality threshold. DeepSeek won the price war. Cursor won the developer-experience war. Anthropic and OpenAI still own the frontier — but the frontier is no longer where the volume is. If you are building on AI in 2026, the math is now brutal: pay frontier prices only when frontier capability is the bottleneck. For everything else, the new mid-tier is good enough. And the composable stack — three tools behind one key — is the workflow that wins. The labs that don't have a credible cheap tier within 60 days will lose volume to DeepSeek and Cursor.