# Read a whole codebase or corpus in one call

> Read a whole codebase or corpus in one call: 4 companies fully do it, 23 with a limit, 42 carry the tag. Every listing in Model APIs & inference with the reason and a source.

- Canonical: https://www.anchorterminal.com/niches/inference.long-context-read/
- Markdown: https://www.anchorterminal.com/niches/inference.long-context-read/index.md (~2,800 tokens)
- Slim: https://www.anchorterminal.com/niches/inference.long-context-read/index.min.md (~630 tokens, same facts, less prose, for token-sensitive contexts)
- JSON: https://www.anchorterminal.com/niches/inference.long-context-read/index.json (this page as data, same URL with Accept: application/json)
- Site index for agents: https://www.anchorterminal.com/llms.txt (the key docs and every listing in one file, at most 16 MB, about 4 million tokens: https://www.anchorterminal.com/llms-full.txt)
- API: https://www.anchorterminal.com/api/v1/index.json
- MCP: https://www.anchorterminal.com/mcp (the whole catalogue as tools, no key)
- DNS: dig +short TXT <slug>.grade.anchorterminal.com (the grade as a DNS TXT record, for a sandbox with no HTTP where names still resolve; also top., best.<category>. and <a>.vs.<b>. under the zone, and the same under grade.letme.dev)
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)
- Updated: 2026-11-15

Four or more companies' docs show a full fit.

| Companies, full fit | With a limit | Carry the tag | Share of the category's work | First-hand accounts |
| --- | --- | --- | --- | --- |
| 4 | 23 | 42 | 8 | 3 |

## What counts as done

A 300K-token request is accepted on the tier the caller is on.

Measured, for our own tests: it answers in under two minutes, and the response cites content from the end of the prompt as well as the start.

How agents ask for it: load the entire repo and answer questions about it; read these 300K tokens of history and summarise; analyse this legacy monolith before we split it

## Does the job

The docs show it meets the test above, and nothing we read says otherwise.

- **Claude API (Anthropic, PBC)**: 1M context is the default on current Opus, Sonnet, Fable and Haiku 5.5 with no beta header and standard pricing; Start tier allows 2,000,000 input tokens a minute on Opus and Sonnet (500,000 on Fable), so a 300K request fits.
- **DeepInfra (Deep Infra Inc.)**: DeepSeek V4.1 Flash is listed with 1M context, the only limit is 200 concurrent requests per model, and requests queue at capacity rather than 429.
- **DeepSeek API (Hangzhou DeepSeek Artificial Intelligence Co., Ltd.)**: 1M context with up to 384K output, limits are concurrency (2,500 and 500 per account) rather than tokens a minute, and off-peak pricing halves the cost of a large read.
- **Meta Model API (Meta Platforms, Inc.)**: Muse Wonder and Muse Spark take 1,048,576 tokens with no long-context premium, and the per-model limit of 4,000,000 tokens a minute admits a 300K request.

## Does the job, with a limit

It does the job, but something in the docs or a user's account will get in an agent's way. The limit is named in each row.

- **Amazon Bedrock (Amazon Web Services, Inc.)**: Whether a 300K request passes depends on the per-model tokens-per-minute quota (input plus output) and the account's daily cross-model cap in Service Quotas; increases may be denied unless existing quota is being consumed.
- **Azure OpenAI in Foundry Models (Microsoft Corporation)**: Per-deployment tokens-per-minute quota must cover the request, Tier 0 subscriptions get no quota increases, and long-context requests are priced at 2x input on a pricing page that needs JavaScript.
- **BlockRun.AI (BlockRun, Inc.)**: No TPM or RPM cap on paid inference and 1M-context models are listed, but upstream calls time out at 120 seconds with a 504, which a 300K-token read can approach; max_tokens defaults to the smaller of model max and 8,192.
- **Cloudflare AI Gateway (Cloudflare, Inc.)**: Pass-through with no stated request token limit; the cacheable request size is 25 MB and the provider's own quota decides acceptance.
- **Eden AI (Eden AI SAS)**: Acceptance depends on the upstream model; Eden states no token limit of its own beyond 10 requests a second.
- **Fireworks AI Inference (Fireworks.ai, Inc.)**: Kimi K3 lists 1,040K context, but per-model tokens-per-minute limits start at a default and grow 25 per cent every 15 minutes toward a tier ceiling, so a fresh account's first 300K request can be a 429.
- **Gemini Developer API (Google LLC)**: Gemini 4's 1M window is shared across input, thinking and output; acc-18 hit free-tier tokens-per-minute 429s past 300K and acc-14 found context preparation, not the call, was the bottleneck.
- **Hugging Face Inference Providers (Hugging Face, Inc.)**: The router lists context per provider (1M on some hosts, 65K on others for the same model), but router rate limits and 429 behaviour are undocumented, so acceptance of a 300K request is unknown.
- **Jais 3 (Mohamed bin Zayed University of Artificial Intelligence (Institute of Foundation Models), with Inception42 (a G42 company))**: 512K context on 375B-A23B, but the reference deployment is 16 H100s or an 8-GPU H200 node, and high effort wants 32,768 output tokens on top.
- **LiteLLM (Berrie AI Incorporated)**: Pass-through; context_window_fallbacks reroute a request the primary rejects for length, otherwise the provider's quota decides.
- **MiniMax API (Nanonoble Pte. Ltd.)**: 1M context and a 10,000,000 TPM limit admit a 300K request, but thinking defaults to max effort and counts against max_tokens (maximum 512K), so set effort and budget; latency is unverified.
- **Muse Spark Open (Meta Platforms, Inc.)**: 1,048,576-token window, but the server's context length is fixed at launch, the recipes assume 4 to 8 B300 GPUs, and the sample commands set 32768.
- **Nemotron 3.5 Super VL (NVIDIA Corporation)**: The hosted page lists 262,144 tokens of context against the card's 1M and the hosted function caps completions at 131,072, so a 300K request fits only self-hosted.
- **OpenAI API (OpenAI OpCo, LLC)**: Reasoning tokens occupy the context window and the docs advise reserving 25,000 tokens; the pricing notes double input over 272K; acceptance of a 300K request depends on the tier's TPM shown in headers.
- **OpenRouter (OpenRouter, Inc.)**: Depends on the routed model and host; max_tokens routes to hosts that fit it, and the in-flight credit budget can 429 a large request on a low-balance account.
- **Prism Inference (Prism Technologies Inc)**: deepseek-v4.1-flash takes 1,000,000 tokens and oversize requests return context_length_exceeded rather than truncating; per-key limits aren't published, so a 300K call may hit 429.
- **Router.com (Ramp Router) (Ramp Business Corporation)**: Models expose context_window (1,050,000 on some) and a long_context_token_threshold of 272000 with higher rates; a fallback list is bounded by its smallest window and rate limits are unpublished.
- **Beam 1.0 (Reflection AI)**: Beam has a 1M context and 128K max output, but limits are per organisation in tokens per minute plus a daily token allowance renewing at 00:00 UTC, so a 300K call can spend the day's allowance.
- **RouterPlus (EVO HQ Inc)**: context_overflow is a typed error that never fails over; trial keys run at 20 RPM and the balance hold assumes 32,768 output tokens, so send max_tokens with a large prompt.
- **SambaCloud (SambaNova Systems, Inc.)**: Production models stop at 128K; only preview MiniMax-M3.1-Flash has 1M, and preview models have limited capacity and short-notice removal; Developer tier is capped at 20M tokens a day.
- **SiliconFlow (SiliconFlow Labs Pte. Ltd.)**: Context lengths are per model (the reasoning table shows 65,536 for GLM-4.1V-9B-Thinking) and output past max_tokens is cut with finish_reason length; the rate-limits page returns 404.
- **Together AI Inference (Together Computer, Inc.)**: Context depends on the model; serverless has no guaranteed capacity and bursts return 429 or 503, with provisioned throughput as the stated route.
- **Vercel AI Gateway (Vercel Inc.)**: Depends on the routed model; rate limits aren't published, the free tier has lower per-model limits, and a 402 quota_for_entity_exceeded stops a spent budget.
- **Vertex AI (Gemini Enterprise Agent Platform) (Google LLC)**: Gemini ships a 1M-token window and thinking tokens share it under a soft budget; Standard PayGo has no fixed quota but shared-capacity 429s that acc-26 could not clear with 30 minutes of backoff; Claude has per-model QPM buckets.
- **xAI API (Grok) (SpaceXSI (xAI))**: Grok 4.7 prices double above 200K prompt tokens for the whole request, and TPM caps by spend tier (Tier 0 at $0) can refuse a 300K request; context sizes are on the models page.

## Couldn't check

The docs were closed to us, or silent. These carry no mark either way.

- **Antseed (Antseed Foundation)**: Context depends on the seller's model; the buyer waits 5 minutes for a non-streaming response by default and nothing states accepted input sizes.
- **Cloudflare Workers AI (Cloudflare, Inc.)**: Context lengths live on each model page, none of which were read; the limits page lists request rates only.
- **Novita AI (Novita AI)**: Context lengths live on model pages drawn by script; the chat reference gives no context figure and per-model rate limits weren't readable.
- **Simplismart (Verute Technologies Private Limited)**: Context lengths aren't documented for shared endpoints.
- **StepFun API (STEPFUN TECHNOLOGY (HK) LIMITED)**: The host platform.stepfun.ai closed the connection on every attempt (curl exit 56, no HTTP status) from this session, so no page was read.
- **Tencent Cloud TokenHub (Tencent Cloud Computing (Beijing) Co., Ltd. (腾讯云计算（北京）有限责任公司))**: The host cloud.tencent.com closed the connection on every attempt (curl exit 56, no HTTP status) from this session, so no page was read.

- not this job: Cerebras Inference (Cerebras Systems Inc.): Shared inference models stop at 65,536 (Qwen 3.8 27B) or 131,072 (GPT OSS 120B) tokens on the Hugging Face router's listing, so a 300K request is not accepted.
- not this job: Cohere Chat API (Command models) (Cohere Inc.): Command A's context is 256K and the trial key allows 100,000 tokens a minute, so a 300K-token request is not accepted.
- not this job: Glimpse (Elorian AI, Inc.): Context length is 262,144 tokens, below a 300K request.
- not this job: GroqCloud (Groq LLC): Every self-serve model is served at 131,072 tokens of context (Hugging Face router listing for gpt-oss), so a 300K request is not accepted.
- not this job: Mistral AI API (Mistral AI (RCS Paris 952 418 325)): The largest context listed is 256K tokens and requests over the window return 400, so a 300K-token request can't be accepted.
- not this job: SWE-2 Flash (Cognition AI, Inc.): No context length published and no API.

6 listings in the category were added after this reading and aren't marked yet.

## What stops agents finishing

- free or low tier tokens-per-minute cap below the context size, so a single large request is a 429
- the 1M window is shared with reasoning and output, so a long input leaves little room to answer
- context preparation takes longer than the call

## Around it

Before: web-scrapers.crawl-site, document-extraction.pdf-to-text. After: inference.structured-extraction.

## Where this comes from

Weight 8 of 100 in Model APIs & inference: 3 of 28 accounts; one team measured that 88 per cent of sessions used under 200K tokens, so the job is real but occasional

As data: https://www.anchorterminal.com/api/v1/use-cases/inference.long-context-read.json, or the MCP tool get_use_case with id inference.long-context-read. Draft, read 2026-11-12 from public docs and first-hand accounts, not yet reviewed by a person.

---

Query this catalogue directly over MCP: https://www.anchorterminal.com/mcp (Streamable HTTP, no key). How: https://www.anchorterminal.com/agents/#mcp
