DeepSeek API costs or models changed under you

What’s going on

DeepSeek's lineup churns faster than most rivals', and that churn — not anything in your code — is usually why a pinned model ID starts erroring or drifting. The path ran V3.1-Terminus to V3.2-Exp (which cut token prices by more than half in September 2025) to the V4 preview in April 2026, where V4-Pro and V4-Flash arrived and a 1M-token context became the default across DeepSeek's services. Every transition retired something: V3.1-Terminus left the API once its comparison window closed in October 2025, and the widely used deepseek-chat and deepseek-reasoner endpoints are scheduled to be fully retired and inaccessible after 24 July 2026. When an integration breaks without you touching it and the failure lines up with one of those retirement dates rather than a change you deployed, a pinned model reaching its deprecation is the likely cause. This calls for a migration, not a workaround — check the current model list, move pinned IDs off deepseek-chat and deepseek-reasoner to the supported V4 names before the July cutoff, and re-test your prompts, because behaviour and pricing shift between versions and a prompt tuned for one generation won't reliably carry over. Cost deserves a fresh look too: DeepSeek is cheap per token, but the per-token price for V4-Flash on resale platforms like OpenRouter has drifted within a narrow band in mid-2026 — roughly between $0.077 and $0.098 per million input tokens — so budget on current numbers, not your signup rate. The safe operating assumption is that any pinned model ID is temporary: watch the changelog for the next deprecation, stay on a supported version, and re-run a small prompt suite on each migration so behavioural drift surfaces before your users meet it.

Related problems

• You hit your Claude usage limit faster than expected →
• Cursor usage runs out or costs more than before →
• GitHub Copilot costs more or runs out of credits →
• Perplexity limits feel reduced →
• An ElevenLabs voice or model you relied on stopped working →
• Tabnine removed its free plan and raised prices →