Changelog

Major changes to AI Foundation Services — new models, feature updates, and deprecations.

v1.2.4 — 2026-11

Aligns the documentation with the Service Description as of 30 October 2026, which is the binding source for everything below. It applies from 1 December 2026; if you use the new models before then, you accept it by using them.

Added

New preview model in the Test catalog: DeepSeek V4.1 Flash (DeepSeek-V4.1-Flash-Preview) from DeepSeek, with a 1M context window, text and image input, streaming, function calling and reasoning. Free on any active plan. See Test / Preview Models.
GLM 5.3 (GLM-5.3) from Z.ai on the Essential, Professional and Agentic plans: built for coding and agentic tasks, with a 1M context window, hosted in Germany. See Plans & Pricing.
Qwen3-Reranker-4B, LightOnOCR-3-4B and PaddleOCR-VL-1.6 moved from preview to production on the Essential, Professional and Agentic plans. The model IDs drop the -Preview suffix: Qwen3-Reranker-4B, LightOnOCR-3-4B, PaddleOCR-VL-1.6. See Reranking and OCR.
New models on the Essential, Professional and Agentic plans: Claude Haiku 5.5 and Gemini 4 Argon. GPT-6 Astra and Claude Fable 5.1 on Professional and Agentic. See Plans & Pricing.
GPT-6 Astra is priced by context length, with separate rates above 272k input tokens; Claude Haiku 5.5 has separate rates above 100k input tokens.
New preview model in the Test catalog for reranking: Qwen3-Reranker-4B (Qwen3-Reranker-4B-Preview), which sorts documents by relevance to a query through the POST /rerank endpoint (compatible with the Cohere and Jina rerank APIs). Free on any active plan. New guide: Reranking, with curl, Python, Node.js and Cohere SDK examples.

Changed

Gemini 3 Flash is no longer scheduled for retirement: the Service Description still lists it, so it stays on the Essential, Professional and Agentic plans.
Qwen3 Embedding 8B costs €0.10 per million input and output tokens (input was €0.15), with new rate limits: 180,000 input TPM and 360 RPM on Essential, 360,000 and 720 on Professional, 600,000 and 1,200 on Agentic.
Enterprise Trust lists the new models, and names the Industrial AI Cloud next to T-Cloud Public for the models Telekom operates.
The per-model examples on Rate Limits show the current limits of GPT-OSS 120B and GPT-6 Luna.

Deprecated

GLM 5.2, GPT-5.6 Luna, Claude Sonnet 5, Claude Opus 5, Claude Fable 5, Gemini 3.1 Pro and Mistral Medium 3 retire on 2026-11-30 and are no longer listed in the plans. They keep working until then. See Scheduled for Retirement for replacements.

Removed

GPT-5, GPT-5 mini, GPT-5 nano, GPT-5 Codex, GPT 5.4, GPT 5.4 mini, GPT 5.5, Claude 4.5 Haiku, Claude 4.6 Sonnet / Opus, Claude Opus 4.8 and Gemini 3 Pro were retired on 2026-10-31 and are removed from the plans. See Retired Models.

v1.2.3 — 2026-10

Aligns the documentation with the Service Description as of 30 September 2026, which is the binding source for everything below.

Added

Responses that contain model output (chat and text completions, the Responses API, images, audio and decisions) now carry the HTTP response header X-AI-Generated-Content: true, which marks the content as AI generated in line with the EU AI Act. The response body is unchanged. See API Endpoints.
Two new preview models in the Test catalog for OCR: LightOnOCR-3-4B (LightOnOCR-3-4B-Preview) from LightOn, Paris, and PaddleOCR-VL-1.6 (PaddleOCR-VL-1.6-Preview). Free on any active plan. New guide: OCR (Document Parsing), with Chat Completions and library examples.
New embedding model on the Essential, Professional and Agentic plans: Qwen3 Embedding 8B (qwen3-embedding-8b), multilingual, with a 32K context window. Same input-token and request limits as Jina Embeddings v2 Base DE; output tokens are not limited, since an embedding returns vectors. See Plans & Pricing.
New API Transition guide. The LLM Serving Service moves to the New API in three phases, with a Legacy endpoint for applications that need the current format; the guide covers the endpoints and what you need to do, API keys and model identifiers, and the changes to requests, responses, streaming and error handling.
New preview model in the Test catalog: d3, a 27B decision model and the recommended model for the Decisions endpoint, with text and image input. Free on any active plan. See Test / Preview Models and the Decisions guide.
Decisions accepts image input on both preview models, d3 and DiffusionGemma 26B A4B. See Image input.
New preview model in the Test catalog: DiffusionGemma 26B A4B, free on any active plan, for chat and for the new Decisions endpoint. See Test / Preview Models and the Decisions guide.
New models on the Essential, Professional and Agentic plans: GPT-6.1 Sol, GPT-6 Luna, Claude Sonnet 5.5, and Gemini 3.8 Flash. Claude Opus 5.5 on Professional and Agentic. See Plans & Pricing.
GPT-6.1 Sol and GPT-6 Luna are priced by context length, with separate rates above 272k input tokens.

Changed

Gemini 3.1 Pro, Gemini 3.5 Flash and Gemini 3 Pro Image now run on Dynamic Shared Quota instead of fixed RPM/TPM limits.
The Service Description shipped with the docs has been updated.

Deprecated

GPT-5 to GPT 5.5, Claude 4.5 Haiku, Claude 4.6 Sonnet / Opus, Claude Opus 4.8 and Gemini 3 Pro retire on 2026-10-31. Gemini 3 Flash retires on 2026-11-30. See Scheduled for Retirement for replacements.

v1.2.2 — 2026-09

Aligns the documentation with the Service Description as of 31 August 2026, which is the binding source for everything below.

Added

GLM 5.3 Flash (GLM-5.3-Flash) and Qwen 3.8 27B FP8 (Qwen3.8-27B-FP8) moved from preview to production on the Essential, Professional and Agentic plans. See Plans & Pricing.
New models on the Essential, Professional and Agentic plans: GPT-5.6 Terra and GPT-5.6 Luna. See Plans & Pricing.
Gemini 3.1 Pro is now priced by context length.
Cached input token pricing on T-Cloud-hosted models. See Prefix Caching.

Changed

The Queue API is not asynchronous. /queue endpoints work like /v2, but when you are over your rate limit they wait up to two minutes instead of returning 429 right away. The guide no longer mentions job polling or background processing. See Queued Requests.
The availability target is 99.5%, not the 99.9% previously published.
GLM 5.2 token prices reduced, and it moved to the standard model tier.
GPT-Image-2 pricing corrected.
Prefix caching is on by default on T-Cloud-hosted models, not opt-in.
Server location and data processing for the Gemini 3 family now read Worldwide instead of Europe. See Enterprise Trust.
The Anthropic models are labelled GCP / Azure.
The Service Description shipped with the docs has been updated.
Enterprise Trust category 1 now lists the models the Service Description names.

Removed

Llama 3.3 70B, Claude 4.5 Sonnet and Claude 4.5 Opus reached their retirement date. See Retired Models for replacements.

Fixed

GLM 5.2 is listed on Coding Models again.
OpenCode: corrected the authentication URL in the plugin instructions.
/category/model-serving now redirects to the guides instead of returning a 404.

v1.2.1 — 2026-08

Added

New preview model in the Test catalog: Qwen 3.8 27B FP8 — natively multimodal, free on any active plan. See Test / Preview Models.

v1.2.0 — 2026-07

Added

New models on the Essential, Professional and Agentic plans: GPT-5.4, GPT-5.4 mini, GPT-5.5 and GPT-Image-2 (OpenAI); Gemini 3.1 Pro and Gemini 3.5 Flash (Google). Claude 4.8 Opus (Anthropic) on Professional and Agentic. See Plans & Pricing.
Mistral Small 4 and Gemma 4 are now generally available, having been preview-only. See Plans & Pricing.
New preview model in the Test catalog: GLM 5.2 — free on any active plan. See Test / Preview Models.

Deprecated

Models across the GPT-4.1, GPT-4o, o-series, Qwen3, Gemini 2.5, Gemini 3 and Claude Sonnet / Opus families are scheduled for retirement on 2026-08-01, along with Mistral Small 3 and GPT Image 1. See Scheduled for Retirement for the table and replacements.

v1.1.0 — 2026-05

Archived snapshot: v1.1.0.

Added

New top-level Models section with a comparison page covering every hosted model — context window, pricing, cloud, modalities, plan availability, and OpenAI-compatible endpoint reference. Image-capable models also surface on the dedicated Vision page via a capability filter.
Changelog page (you are reading it).
Two new preview models in the Test catalog: Mistral Small 4 (Mistral-Small-4-119B-2603-Preview) and Gemma 4 (gemma-4-31B-it-FP8-preview). See rate limits for the preview-model policy.

Removed

Fine-Tuning API documentation. Requests to /guides/fine-tuning now redirect to the optional services. Superseded: fine-tuning is offered as an optional service on request — see § 3.4.2 of the Service Description. The self-service API documentation remains withdrawn.

v1.0.0

Archived snapshot: v1.0.0 — the initial released state of the docs prior to the changelog being introduced.