Changelog — pools, models & platform updates

Every public change to our model pools and platform, newest first — models added or retired, pool launches, and new API capabilities. For the current lineup and prices, see Plans & Limits.

September 2026

2026-09-10 — New model: DeepSeek V4.1 Flash

DeepSeek’s new-generation V4.1 Flash is now served in the Core Pool as deepseek-v4.1-flash, replacing V4 Flash in place — a new architecture with native image input and a 1M-token context, on the same subscriptions. Every Core subscriber has it now with nothing to change on existing keys. See DeepSeek V4.1 Flash.

2026-09-10 — Model retired: DeepSeek V4 Flash

DeepSeek V4 Flash has been upgraded to DeepSeek V4.1 Flash. Requests with model: "deepseek-v4-flash" are served by V4.1 Flash until 2026-10-10, then return an invalid-model error — migration is one line: see DeepSeek V4 Flash — retired.

August 2026

2026-08-30 — New model: GLM 5.3

Z.ai’s GLM 5.3 is now served in the Frontier Pool as glm-5.3, replacing GLM 5.2 in place — same base model with extended post-training, on the same subscriptions; every Frontier subscriber has it now with nothing to change on existing keys. See GLM 5.3.

2026-08-30 — Model retired: GLM 5.2

GLM 5.2 has been upgraded to GLM 5.3. Requests with model: "glm-5.2" are served by GLM 5.3 until 2026-09-30, then return an invalid-model error — migration is one line: see GLM 5.2 — retired.

2026-08-15 — Model retired: Kimi K2.7

Kimi K2.7 has been retired from the Frontier Pool, which continues with GLM 5.2 and MiniMax M3 on the same subscriptions. For the Kimi family, Kimi K3 — Moonshot’s current flagship — is served in the Flagship Pool. Migration is one line: see Kimi K2.7 — retired.

2026-08-14 — New model: Qwen3.8 Max

Alibaba’s flagship Qwen3.8 Max joins the Flagship Pool next to Kimi K3 — 1M-token context, vision input, hybrid per-request reasoning, on the same unlimited time-block subscriptions. Every Flagship subscriber can use it now with model id qwen3.8-max — nothing to change on existing keys. See Qwen3.8 Max.

July 2026

2026-07-31 — Model upgraded: DeepSeek V4 Flash (0731 build)

DeepSeek released V4-Flash-0731, a retrain of V4 Flash focused on agent work, and the Core Pool now serves it. Same model id (deepseek-v4-flash), same 1M-token context — nothing to change on your side. See DeepSeek V4 Flash.

2026-07-28 — New pool: Flagship

The Flagship Pool launches with Moonshot’s Kimi K3 as its sole model, on the same unlimited time-block subscriptions as the other pools. Seats are very limited. See Kimi K3 and Plans & Limits.

2026-07-27 — Model retired: Kimi K2.6

Kimi K2.6 has been retired from the Frontier Pool. Kimi K2.7 — the newer model in the same family — remains available to every Frontier subscriber.

2026-07-25 — Platform: reasoning control

Reasoning (extended thinking) can now be enabled or disabled per request, with the same control on both the OpenAI- and Anthropic-compatible endpoints. See Reasoning.

2026-07-09 — Platform: combined keys

One API key can now span several of your Unlimited subscriptions: coverage windows add up, and overlapping hours stack your parallel capacity. See Combined keys.

2026-07-06 — Platform: MCP server

The catalog and subscription flow are now exposed as a remote MCP server — browse pools, get a live USDC quote, subscribe, and retrieve connection snippets from any MCP client. See MCP Server.

2026-07-05 — New pools: Frontier and Core

The catalog is now organized in two pools, each sold as an unlimited time-block subscription that covers every model in the pool: the Frontier Pool (Kimi K2.7, Kimi K2.6, GLM 5.2, MiniMax M3) and the Core Pool, which debuts with two new models — DeepSeek V4 Flash and MiMo v2.5.

2026-07-03 — New model: Kimi K2.7

Kimi K2.7, the successor to Kimi K2.6, joins the catalog. See Kimi K2.7.

June 2026

2026-06-19 — New model: MiniMax M3

MiniMax M3 replaces MiniMax M2.5 in the catalog. See MiniMax M3.

2026-06-17 — New model: GLM 5.2

GLM 5.2 replaces the retired GLM 4.7. See GLM 5.2.

2026-06-02 — New model: Kimi K2.6

Moonshot’s Kimi K2.6 joins the catalog. See Kimi K2.6.

Earlier in 2026

February 2026 — Platform: x402 payments and Anthropic-compatible endpoint

From launch, keyless API requests are answered with HTTP 402 and USDC pricing on Base L2 via the x402 protocol, and the Anthropic Messages endpoint lets the Anthropic SDK and Claude Code connect by changing only the base URL.