Morph
Listed at https://morphllm.com
Overall Rank #41 ⭐ Consider
❌ No mainland China direct endpoint. Base URL https://api.morphllm.com is a single global endpoint, hosted by AutoInfra, Inc. (YC-backed US company). OpenRouter routes morph/morph-v3-fast and morph/morph-v3-large through multi-region, but the official Morph API is not available from mainland China without a proxy. China production traffic sees typical 200-400 ms transpacific first-byte latency. | 🌍 International
💰 Token Pricing
| Type | Price | Note |
|---|---|---|
| Input | 按模型分别计费，按每 1M 输入 token 计（文本模型）。从 llms.txt 验证（2026-08-28）：Kimi K3 $2.90、Qwen 3.5 397B $0.50、GLM-5.2 744B $1.10、GLM-5.3-Flash $0.15、MiniMax M3 $0.30、MiniMax M2.7 $0.279、DeepSeek V4 Flash (beta) $0.12、Qwen 3.8/3.6 27B $0.289、Gemma 4 31B $0.14。Qwen 3.5 397B 缓存输入 $0.30 per 1M（其他模型无独立缓存价）。Fast Apply 用 `auto` 模型按 prompt+completion 两档计（OpenRouter 验证 morph-v3-fast $0.0008/$0.0012 per 1k、morph-v3-large $0.0009/$0.0019 per 1k），Compact、Reflexes、WarpGrep 与 Router 按事件计费（详见 pricingEN）。 | per million tokens |
| Output | 按模型分别计费，按每 1M 输出 token 计（文本模型）。从 llms.txt 验证：Kimi K3 $14.00、Qwen 3.5 397B $3.50、GLM-5.2 744B $4.10、GLM-5.3-Flash $0.50、MiniMax M3 $1.20、MiniMax M2.7 $1.20、DeepSeek V4 Flash (beta) $0.278、Qwen 3.8/3.6 27B $2.40、Gemma 4 31B $0.40。Fast Apply 输出按 morph-v3-fast $1.20 / morph-v3-large $1.90 per 1M token；Fast Apply 用 `auto` 推荐档（5,000-10,500 tok/s, ~98% 精度）按实际路由模型分档计费。Kimi K3 `service_tier: standby` 走 GLM-5.2 同等待遇，标准按 token 计。 | per million tokens |
| Cache Read | Qwen 3.5 397B 缓存输入折扣价 $0.30 per 1M（折扣 40%）；其他开放权重模型无独立缓存价（llms.txt: 'the other models have no separate cache rate'）。Fast Apply / Compact / Reflexes / WarpGrep / Router 按事件或按 token 计费，无缓存档。 | Discounted |
💡 Free Credits: No permanent free tier (no zero-cost starter plan). New signups get an API key and are billed pay-as-you-go from the first token. Morph offers up to $5,000 in Startup Credits for qualified startups, applied via a separate sales contact form (morphllm.com home page). The homepage CTA 'Start free. Get API Key' refers to signup friction, not to a free quota.
🤖 Supported Models (13)
Kimi K3 (2.8T MoE) — flagship Fable-tier open-weight model at ~100 tok/s with 1M context; per-1M input/output token pricing (verified on docs.morphllm.com llms.txt Fast Models table 2026-08-28)GLM-5.2 (744B MoE) — Opus-tier open-weight model with 1M context; OpenAI `service_tier` parameter supported for default/standby processingQwen 3.5 397B — open-weight dense model at ~180 tok/s with 262k context; cached input bills at $0.30 per 1MGLM-5.3-Flash — fast 1M-context tier priced for high-volume agent workloadsDeepSeek V4 Flash (beta) — 1M-context model with sub-second first-tokenQwen 3.8 27B / Qwen 3.6 27B — dense low-latency models tuned for coding agents at ~100 tok/s with 131k contextMiniMax M3 428B / MiniMax M2.7 230B — 256k/196k context MoE for agentic workflows at ~90 tok/sGemma 4 31B — open-weight dense at ~120 tok/s with 175k contextFast Apply (morph-v3-fast 10,500+ tok/s / morph-v3-large 5,000+ tok/s) — specialized code-merge model with 96-98% accuracy on editsCompact (morph-compactor) — verbatim context compression at 33,000 tok/s; 100k tokens compresses in under 2 secondsWarpGrep (morph-warp-grep-v2.1) — semantic code search subagent with built-in tools (grep_search, read, list_directory, glob, finish)Reflexes (jailbreak, guardrail, leaked-thinking, stuck-in-a-loop, user-frustrated, etc.) — text classifiers running at ~90 ms per labelModel Router — classifies prompt difficulty/ambiguity/domain in ~180 ms; $0.005 per request
✨ Pros
- ✓Inference built for agents: Fast Apply / Compact / WarpGrep / Reflexes / Model Router cover the full agent loop (edit → search → compress → monitor → route) on one stack
- ✓OpenAI + Anthropic wire formats: a single API key serves both /v1/chat/completions and /v1/messages, so the official OpenAI and Anthropic SDKs work — Claude Code and Codex run out of the box
- ✓Open-weight flagship lineup: Kimi K3 (2.8T), GLM-5.2 (744B MoE), Qwen 3.5 397B, DeepSeek V4 Flash, MiniMax M3/M2.7, Gemma 4 31B, Qwen 3.8/3.6 27B
- ✓Fast Apply speed: morph-v3-fast at 10,500 tok/s with 96% accuracy; morph-v3-large at 5,000 tok/s with 98% accuracy; the `auto` route picks the right one
- ✓Compact at 33,000 tok/s: 100k tokens compress in under 2 seconds, with byte-for-byte preservation of any `<keepContext>` tagged content
- ✓WarpGrep semantic code search subagent: ~6 seconds per query, with built-in grep_search / read / list_directory / glob / finish tools (no tool definition required from the caller)
- ✓Reflexes: 11 built-in classifiers (jailbreak / guardrail / leaked-thinking / stuck-in-a-loop / ambiguity / difficulty / domain / etc.) returning labels in ~90 ms
- ✓Claude Code / Codex / Cursor integration: the MCP server `npx @morphllm/morphmcp@latest` exposes edit_file / codebase_search / github_codebase_search as native tools
- ✓Y Combinator-backed: Y Combinator logo and 'Backed by Y Combinator' appear on morphllm.com footer; OpenRouter Top 25 provider; 100B+ tokens/day, 400+ production agents
⚠️ Cons
- ×No China-region endpoint: the base URL https://api.morphllm.com is single-endpoint with no Asia-Pacific or China route, so China production traffic sees 200-400 ms transpacific latency
- ×No monthly subscription / no permanent free tier: pure pay-as-you-go plus separately-applied Startup Credits (up to $5K) — evaluation cycles must self-fund
- ×Pricing skews toward high-volume agent workloads: flagship Kimi K3 2.8T at $2.90/$14.00 per 1M is well above general-purpose chat tiers like GLM-5.3-Flash ($0.15/$0.50)
- ×Fast Apply requires a proprietary prompt template `<instruction>{...}</instruction><code>{...}</code><update>{...}</update>` to return the merged file — not a drop-in OpenAI replacement
- ×WarpGrep is a multi-turn tool-calling loop: the API returns tool_calls and the developer must execute grep_search / read / list_directory / glob locally on the filesystem (typically 3-6 turns) before finish
- ×GLM-5.2 `service_tier: standby` returns 429 when fleet utilization exceeds ~25%, with no SLA — not suitable for latency-sensitive interactive traffic
- ×Compact / WarpGrep / Reflexes / Router are OpenAI-format only — no Anthropic Messages wire format (only chat models serve both)
- ×Newer entrant (YC-stage AutoInfra, Inc.): community, third-party SDKs, and tutorials lag well behind OpenAI / Anthropic / DeepSeek
- ×No affiliate program: no partner revenue share (cf. Cloudflare / OpenRouter)
🎯 Best For
Teams building production coding agents or long-running agents who need the full Fast Apply (10,500 tok/s code merge) + Compact (33,000 tok/s context compression) + WarpGrep (~6s semantic code search) + Reflexes (jailbreak / stuck-in-loop monitor) + Model Router stack on one API; teams migrating Claude Code / Codex / Cursor onto open-weight models (Kimi K3, GLM-5.2, Qwen 3.5 397B) via the MCP server `npx @morphllm/morphmcp@latest`; workloads that need Anthropic Messages wire format on open-weight inference; startups eligible for up to $5K in Morph Startup Credits.
💰 Pricing & Plans
| Product / Model | Billing Unit | Input / In | Output / Out | Notes |
|---|---|---|---|---|
| Kimi K3 2.8T (morph-kimik3) | Per 1M tokens | $2.90 | $14.00 | Flagship Fable-tier; 1M context; ~100 tok/s |
| Kimi K3 Fast (morph-kimik3-fast) | Per 1M tokens | $6.00 | $22.50 | Fastest Kimi K3 tier; same 1M context |
| Qwen 3.5 397B (morph-qwen35-397b) | Per 1M tokens | $0.50 ($0.30 cached) | $3.50 | 262k context; ~180 tok/s; cached input 40% off |
| GLM-5.2 744B (morph-glm52-744b) | Per 1M tokens | $1.10 | $4.10 | 1M context; ~80 tok/s; supports service_tier standby |
| GLM-5.3-Flash (morph-glm53flash) | Per 1M tokens | $0.15 | $0.50 | 1M context; budget tier for high-volume agents |
| MiniMax M3 428B (morph-minimax3-428b) | Per 1M tokens | $0.30 | $1.20 | 256k context; ~90 tok/s MoE |
| MiniMax M2.7 230B (morph-minimax27-230b) | Per 1M tokens | $0.279 | $1.20 | 196k context; ~90 tok/s MoE |
| DeepSeek V4 Flash beta (morph-dsv4flash) | Per 1M tokens | $0.12 | $0.278 | 1M context; ~150 tok/s; sub-second TTFT |
| Qwen 3.8 / 3.6 27B | Per 1M tokens | $0.289 | $2.40 | 131k context; ~100 tok/s dense |
| Gemma 4 31B (morph-gemma4-31b) | Per 1M tokens | $0.14 | $0.40 | 175k context; ~120 tok/s dense |
| Fast Apply (morph-v3-fast) | Per 1k tokens (OR) | $0.0008 | $0.0012 | 10,500 tok/s; 96% edit accuracy; `auto` recommended |
| Fast Apply (morph-v3-large) | Per 1k tokens (OR) | $0.0009 | $0.0019 | 5,000 tok/s; 98% edit accuracy |
| Compact (morph-compactor) | Per event / per token | n/a | n/a | 33,000 tok/s; 100k tokens < 2s; preserve <keepContext> byte-for-byte |
| WarpGrep (morph-warp-grep-v2.1) | Per request | n/a | n/a | ~6s per query; built-in grep/read/list/glob/finish tools |
| Reflexes (realtime) | Per event | $0.001 → $0.0005 | n/a | 11 built-in classifiers; ~90 ms per label; 1M-event tiered discount |
| Reflexes (async batch) | Per event | $0.0005 → $0.00025 | n/a | Same classifiers, batch mode |
| Model Router | Per request | $0.005 | n/a | Prompt difficulty/ambiguity/domain classifier; ~180 ms |
| Reserved B200 | Per GPU-hour | Contact sales | Contact sales | Dedicated capacity for production agent fleets |
🔧 API & Developer Experience
- •API Style: OpenAI-compatible REST at https://api.morphllm.com/v1 — the official OpenAI Python/TypeScript SDK works as a drop-in by changing base_url. Every open-weight chat model also serves the Anthropic Messages wire format at /v1/messages (with /v1/messages/count_tokens), so the official Anthropic SDKs and Claude Code / Codex run unchanged.
- •Endpoints: /v1/chat/completions (all chat models + Fast Apply via model=morph-v3-fast|auto), /v1/messages (Anthropic format for chat models), /v1/messages/count_tokens, /v1/compact (native Morph Compact), /v1/responses (OpenAI Responses API for Compact), /v1/reflex/predict (Reflexes), /v1/router/classify + /v1/router/multimodel (Model Router), /v1/fine_tuning/* (custom Reflexes), /v1/models (model list with pricing and context length).
- •Auth: Bearer token (single MORPH_API_KEY) used for every endpoint; Anthropic-format calls also accept `x-api-key: KEY` header. API keys are created in the Morph dashboard at morphllm.com/dashboard/api-keys and apply to all products — Apply, Compact, WarpGrep, Router, Fast Models, Custom Reflexes.
- •Fast Apply Prompt Format: Proprietary `<instruction>{brief description}</instruction><code>{original full file}</code><update>{partial edit with // ... existing code ... markers}</update>` template. The model returns the fully merged file in `choices[0].message.content` (or Anthropic `content[-1].text`). Use the `auto` route to let Morph pick between morph-v3-fast (10,500 tok/s, 96%) and morph-v3-large (5,000 tok/s, 98%) based on edit complexity.
- •Compact & keepContext Tags: POST /v1/compact takes `{ input, query?, compression_ratio, preserve_recent }` and returns `{ output, compacted_line_ranges }` so the caller can see exactly what survived. Wrapping sections in `<keepContext>` guarantees byte-for-byte preservation regardless of compression ratio. Also accessible via `model: morph-compactor` on /v1/chat/completions and via /v1/responses.
- •WarpGrep Tool Loop: morph-warp-grep-v2.1 emits `tool_calls` for grep_search / read / list_directory / glob / finish. The developer must execute each tool call on the local filesystem and reply with `{role: 'tool', tool_call_id, content}` messages. Typically 3-6 turns. The model handles path resolution and snippet extraction; the developer handles I/O.
- •MCP Server: `npx @morphllm/morphmcp@latest` exposes `edit_file`, `codebase_search`, and `github_codebase_search` as native MCP tools for Claude Code / Cursor / Windsurf. `claude mcp add morph -- npx -y @morphllm/morphmcp@latest` registers it; MORPH_API_KEY in the env covers the auth.
- •Claude Code & Codex Integration: Four env vars route Claude Code through Morph: `ANTHROPIC_BASE_URL=https://api.morphllm.com`, `ANTHROPIC_AUTH_TOKEN=MORPH_API_KEY`, `ANTHROPIC_MODEL=morph-kimik3`, `ANTHROPIC_SMALL_FAST_MODEL=morph-kimik3-fast`. Sonnet/Opus/Haiku map to Morph models via the base URL; File edits, Bash, Grep tools still work through Anthropic's tool_use loop.
- •Model Router: /v1/router/classify returns `{ difficulty, ambiguity, domain }` labels with confidence and `meets_threshold`. /v1/router/multimodel takes `allowed_models` / `allowed_providers` / `policy: balanced|cost_efficient` / `default_model` and returns the single best model to call next (e.g. claude-haiku-4-5-20251001 for an easy prompt at 0.93 confidence).
- •Service Tiers (GLM-5.2): OpenAI's `service_tier` parameter maps to GLM-5.2: `default`/`auto`/omitted = standard processing; `standby` = best-effort capacity with no SLA, returns 429 with `Retry-After` when fleet is over ~25% utilized (nothing generated, nothing billed). Retry with backoff or send on `default` for interactive traffic.
- •SDKs: Official TypeScript SDK `@morphllm/morphsdk` (npm), official Python SDK `morphsdk` (pip), MCP server `npx @morphllm/morphmcp@latest`. Anthropic SDK and OpenAI SDK both work natively via base_url change. Fast Apply / WarpGrep / Compact / Reflexes / Router are reached via raw HTTP or the MorphClient wrapper.
🧠 Inference Built For Agents
Morph's positioning is the agent loop, not the chat. A coding agent spends most of its tokens on four operations: searching a repo for the right snippet, applying an edit to a file, compressing context so the next LLM call still fits, and routing easy vs. hard prompts to the cheapest model that handles them. Morph ships a specialized inference endpoint for each of these — WarpGrep for sub-6s semantic search with built-in tools, Fast Apply at 10,500 tok/s with byte-for-byte merged output, Compact at 33,000 tok/s with `<keepContext>` byte preservation, and the Model Router that classifies prompt difficulty in ~180 ms — all behind one API key and one OpenAI-compatible base URL. The open-weight chat lineup (Kimi K3 2.8T, GLM-5.2 744B, Qwen 3.5 397B, DeepSeek V4 Flash, MiniMax M3/M2.7, Qwen 3.8/3.6 27B, Gemma 4 31B) runs through both the OpenAI and Anthropic wire formats, letting Claude Code, Codex, Cursor, and any other agent switch from Anthropic to open-weight models with four environment variables. Y Combinator-backed AutoInfra, Inc. serves 100B+ tokens per day across 400+ production agents and ranks in the OpenRouter Top 25.
🌐 Regional Availability & Latency
Morph runs a single global endpoint at https://api.morphllm.com with no documented regional or Asia-Pacific routing layer. Production traffic from North America and Europe sees low single-digit tens of ms first-byte latency; mainland China requires a proxy or transit, with typical 200-400 ms transpacific first-byte latency. The OpenRouter-published Morph models (morph/morph-v3-fast, morph/morph-v3-large) are routed through OpenRouter's own multi-region infrastructure, which gives Chinese developers a low-latency Fast Apply path without leaving the OpenRouter API key — but the open-weight chat lineup (Kimi K3, GLM-5.2, Qwen 3.5 397B, etc.) is only on Morph's direct API and OpenAI/Anthropic direct, not through any region-aware proxy. No ICP-licensed mainland China endpoint as of August 2026.