Changelog

Product updates and platform changes from the sference team.

2026-11-16

Off-peak pricing

  • PricingEvery chat-model request that arrives between 22:00 and 06:00 Europe/Berlin (21:00 to 05:00 UTC in winter, 20:00 to 04:00 UTC in summer) bills at 50% off. Nothing to enable: same endpoints, same models, realtime latency. Arrival time decides the price, so each call of an agent run is priced on its own.
  • PricingOff-peak combines with the tiers: realtime and flex both bill 0.5× list at night (flex takes the lower factor, they don't stack), priority bills 0.65× list, and Batch API jobs stay at 0.5× around the clock. Decisions models and custom models are not discounted.
  • APIGET /v1/models publishes the window per model as pricing.off_peak (start, end, timezone, price_factor), so a scheduler can read it instead of hardcoding hours. Off-peak requests are flagged in the console's activity view. Details at /docs/pricing#off-peak-pricing.

2026-11-11

DeepSeek V4.1 Flash price cut

  • Pricingdeepseek-ai/DeepSeek-V4.1-Flash-1017 drops to $0.35 / $1.32 per 1M tokens ($0.03 cached), from $0.50 / $1.50 ($0.05 cached), effective 11 November at 18:00 CET. The undated alias bills at the same rate.

2026-10-18

Idempotency keys on create endpoints

  • APIPOST /v1/chat/completions, /v1/responses, /v1/messages, and /v1/batches now accept an Idempotency-Key header (1–256 characters). A retry with the same key and the same body returns the original resource instead of creating and billing a second one; the same key with a different body returns 409 idempotency_key_reused. Keys are scoped to the organization, API key, and creator, and deduplicate for 24 hours — the same semantics as OpenAI's Idempotency-Key.
  • PlatformReplays render the stored result rather than re-running the request: streaming retries replay the original SSE events, background responses can be retrieved while still running, and a request cancelled before it produced a billable result re-executes on retry instead of replaying the cancellation.
  • DocsThe new idempotency guide at /docs/guides/idempotency covers retry patterns, conflict handling, and the 24-hour window.

2026-10-17

DeepSeek V4.1 Flash weights refresh, structured-output fixes

  • Modelsdeepseek-ai/DeepSeek-V4.1-Flash-1017 is the new canonical id for DeepSeek V4.1 Flash: DeepSeek's October 17 weights refresh, served at the same $0.50 / $1.50 per 1M tokens ($0.05 cached) with the same 1M context, tool calls, and image input. The previous undated id now resolves to it as an alias and is deprecated — existing requests keep working, but pin the dated id for new work.
  • PlatformA Claude Code built-in tool schema — a string pattern combined with minLength/maxLength, which is legal JSON Schema — was rejected by the structured-output ingress lint, so every Claude Code request routed to a sference model failed with a 400. Ingress now normalizes the combination to what the grammar engine actually enforces (the pattern is kept, the length bounds are dropped) and lets the request through, with each rewrite counted for observability.
  • APIStructured-output schemas the grammar engine genuinely cannot compile (if/then/else branches, multipleOf) are now rejected at intake with an actionable 400 naming the unsupported construct, instead of failing in the engine as a 500.
  • APISync requests without an explicit output cap now default to 16k max output tokens, and a client cap that cannot fit the sync wait window is rejected with a 400 up front instead of timing out with a 504.

2026-10-07

Model cards

  • DocsEvery live model now has a card at /models/<catalog-id>: pricing, context and capabilities, how it fits the portfolio, a quickstart, and sourced benchmarks. Generated from the live catalog, with a Markdown twin for agents.

2026-10-06

Clef 27B: realtime decisions

  • ModelsCloudflare's open-weight 27B decision model, Cloudflare/clef, is serving on European GPUs. Send one input and up to 64 typed questions to POST /v1/decisions and get calibrated probabilities for every option back. $0.24 per 1M input tokens, output free, 64K context.
  • ModelsIndependently benchmarked by Requesty on 5 October over 22,000 decisions: 97.3% of answers match Clef served on Cloudflare, 94.3% on Banking77, and 0.02–0.04 calibration error.
  • PlatformQueue pressure now fails fast with 429 instead of timing out, and client disconnects report as 499 rather than 500.
  • PlatformImage-heavy request floods can no longer degrade the API.
  • PlatformFixed prefill head-of-line blocking in SGLang, where one cold 600K-token prompt could hold every other request's first token for 105 s. The fix is upstreamed.

2026-09-22

Sub-processor list update

  • LegalAdded Daytona (GPU compute in France), Attio, Linear, Slack, Notion, and GitHub to the published sub-processor list. None of the non-compute vendors receive inference content.

2026-09-14

DeepSeek V4.1 Flash, catalog changes

  • Modelsdeepseek-ai/DeepSeek-V4.1-Flash is live at $0.50 / $1.50 per 1M tokens ($0.05 cached): 1M context, tool calls, streaming, and image input, new on the DeepSeek line. V4 Flash keeps running; retirement will be announced with a date.
  • ModelsQwen3.6 35B, Qwen3-VL 30B, and ThinkingCap Qwen3.6 27B are no longer served. GLM 5.3 Flash covers both text and vision at $0.20 / $0.60 with a 1M window.

2026-09-10

Agent-readable docs

  • DocsEvery docs page, API-reference page, blog post, and legal page is available as Markdown: append .md to the URL, or send Accept: text/markdown.

2026-09-07

Pricing update, GLM 5.3, status page

  • PricingDeepSeek V4 Flash moves to $0.28 / $0.56, with cached input held at $0.07. Contracted and dedicated rates are unaffected.
  • ModelsGLM 5.3 replaces GLM 5.2 at the same $1.20 / $4.20 ($0.26 cached) and the same 1M context. GLM 5.3 Flash joins at $0.20 / $0.60 with image input.
  • PlatformPublic status page with incident notifications at status.sference.com.

2026-08-20

Security & compliance hardening

  • SecurityEnabled Amazon GuardDuty threat detection with findings forwarded to on-call, and multi-region CloudTrail audit logging.
  • SecurityVPC Flow Logs enabled on all networks; log retention standardized to 365 days.
  • SecurityExtended monitoring: load-balancer health/latency, database, and server alarm coverage.
  • PlatformDatabase audit logging (pgaudit) configured for the production Postgres fleet.

2026-08-12

Anthropic Messages API

  • APIPOST /v1/messages speaks the Anthropic Messages API, so the anthropic SDKs and Claude Code run on open models by changing the base URL. Streaming, tool use, extended thinking, and vision are supported.

2026-08-10

Encryption & access controls

  • SecurityAll S3 storage now enforces HTTPS-only access via bucket policy.
  • SecurityEmployee AWS access consolidated to SSO with MFA; standing console accounts removed.
  • PlatformPoint-in-time recovery enabled for infrastructure state.
  • PlatformBranch protection now enforced on all repositories: required reviews and status checks.

2026-08-05

DeepSeek V4 Flash 0731 release

  • Modelsdeepseek-ai/DeepSeek-V4-Flash now serves DeepSeek's updated 0731 release (also addressable as deepseek-ai/DeepSeek-V4-Flash-0731), evaluated against the previous build before the switch. Pricing and the 1M context window are unchanged.

2026-08-04

Dependency security updates

  • SecurityPatched high-severity vulnerabilities in cryptography (CVE-2026-69247) and httplib2 (CVE-2026-59939) — same-day turnaround.