每一次重大 AI 模型发布。追踪、验证、解释。

Z.ai's 14 Aug 2026 release reuses GLM-5.2's base model — every gain comes from scaled post-training on long-horizon RL environments. Large jumps on agentic coding (Terminal-Bench 3.0 4.6 → 28.3, SWE-Marathon 19.4 → 42.5) and an unplanned cyber result (CyberGym 84.5). All figures are Z.ai's own; no independent evaluation exists yet. Weights promised two weeks after launch.

为何重要: It reaches Opus 4.8's coding score on roughly 2.4× fewer output tokens, so the saving is on the bill rather than the leaderboard — but check your API calls first: `thinking.type: "disabled"` is no longer supported and will fail on glm-5.3.

xAI's 12 Aug 2026 successor to Grok 4.5: same $2/$6 API price, same 500K context, a longer post-training run aimed at long-running agents and visual/interactive work. A faster variant costs double. Knowledge cut-off is 1 Feb 2026.

为何重要: Frontier-class intelligence at the old 4.5 price — migrate off 4.5 with a one-line model swap. It is a knowledge-work pick, not the cheapest volume model and not the strongest terminal coding agent.

The deepseek-v4-pro API id now serves DeepSeek-V4-Pro-0813 weights. Calling method and promo pricing are unchanged. DeepSeek's changelog still has no Pro GA entry — the last note (31 Jul) said an official Pro release would follow soon — so this is a weight drop, not a new product.

为何重要: Harder reasoning stays cheap on the same endpoint. Keep Flash ($0.14/$0.28) for volume; do not treat 0813 as a priced relaunch until DeepSeek posts a changelog.

ByteDance's new video model generates 30-second audio-and-video clips in one pass, extends them over multiple rounds for multi-minute pieces, and edits by timestamp. At launch it runs inside Jimeng AI and Doubao Pro, with API access announced as coming via BytePlus ModelArk rather than available.

为何重要: The clearest jump yet in one-take video, but there is no general API to wire into a workflow — treat it as something to try inside a consumer video app, not a tool to rebuild your content process around until the API actually ships.

The official V4 Flash checkpoint entered public beta on the existing deepseek-v4-flash API id, replacing the April preview. The legacy deepseek-chat and deepseek-reasoner ids were retired on 2026-07-24.

为何重要: Near-frontier results at roughly a tenth of typical API prices, with MIT-licensed open weights — the budget and privacy-sensitive pick just got better without changing price.

Anthropic's new everyday flagship: close to Fable 5 on most benchmarks at half the price ($5/$25 vs $10/$50 per MTok), with a low/medium/high effort toggle to trade cost against capability.

为何重要: Frontier-level output at half the flagship price — re-check which Claude tier your workflows actually need before renewing.

Moonshot's 2.8-trillion-parameter sparse MoE — reportedly the largest open-weight model yet, with a 1M-token context window. Weights announced for late July; until then benchmark claims are vendor-reported.

为何重要: Open-weight models keep closing on the paid frontier — if you pay per-token for API work, the cheap tier just got stronger again.

OpenAI's GPT-5.6 family in three tiers — Sol (frontier), Terra (balanced), Luna (fast/cheap) — general availability across ChatGPT, Codex, and the API on Jul 9 after a government-vetted limited preview in late June.

为何重要: Three clear price tiers make it easier to match the model to the job — most small-business tasks belong on the cheap tier, not the flagship.

xAI's first model built specifically for coding and agentic work, priced aggressively under Anthropic and OpenAI flagships with a 500K context window.

为何重要: Agentic coding on a budget is now a three-way price war — worth re-testing your coding stack before renewing anything.

A Mythos-class model made safe for general use — Anthropic's most capable generally available model, sitting above the Opus tier.

为何重要: The frontier of general-purpose reasoning just moved again; capable assistants keep getting cheaper to match.

The Mythos-class model available to approved organizations without the general-use safety measures applied to Fable 5.

为何重要: Signals how fast the top tier is advancing — the same capability reaches everyone shortly after.

A strong all-round released model, widely cited as a top performer through mid-2026.

为何重要: A dependable default for hard reasoning, coding, and long-document work.

OpenAI's mid-2026 frontier update, trading the top spot with Claude Opus on many benchmarks.

为何重要: Keeps the price-for-capability race moving — good news for anyone paying per token.

A March 2026 frontier release with a 1M-token context window and strong computer-use scores.

为何重要: Million-token context means it can read whole manuals, contracts, or codebases at once.

An open-weights frontier model with a 1M+ token context window and strong coding scores.

为何重要: Open weights + very low cost make it the value pick for budget-conscious and privacy-sensitive setups.

Google's February 2026 frontier update to the Gemini 3 line, with a very large context window.

为何重要: Deep integration with Google Workspace makes it a natural fit if you live in Docs and Gmail.

A February 2026 iteration in the GPT-5 line ahead of the March 5.4 release.

为何重要: Part of the steady cadence keeping the mainstream assistant sharp.

A February 2026 Opus update (alongside Sonnet 4.6), continuing Anthropic's rapid iteration.

为何重要: Reliability gains at the same price point — worth re-testing your prompts on each bump.

The Gemini 3 flagship that opened the current generation for Google.

为何重要: Set the bar for long-context multimodal work heading into 2026.

A late-2025 Opus release that anchored Anthropic's top tier into 2026.

为何重要: The baseline many businesses standardized on before the 2026 wave.