Beam 1.0

# Fix list: Beam 1.0

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/reflection-beam, the October 2026 research run, assessed 17 October 2026. Grade D, 53.4 out of 100.

This is everything the published grade says the listing lacks, the biggest possible gain to the total first. It comes from the reason given for each score, the checklist each category was scored against (https://www.anchorterminal.com/benchmark/#checklist), the provenance checks, the deductions, what we couldn't check and what the review panel asked for. A fix counts at the next check, once it's public.

For a coding agent working on Beam 1.0: work through the items below in the product, its docs and its public pages. Each category gives the reason for its score, with the points each checklist item earned, and the checklist itself, so the gap is the items that earned less than their points. Change the product, not the wording, and keep a note of what you changed and where it's published.

## 1. Reliability, 40 out of 100, up to 12 more on the total

Why it scored 40: Hosted lines, for the Reflection API. Status page at status.reflection.ai on incident.io, with API, platform and docs components and an RSS feed (20). Its history from September 2026 holds three API incidents, a serving cluster down from 22:01 to 23:20 UTC on 14 October (critical, about 80 minutes), elevated errors from 06:10 to 07:10 UTC on 15 October (critical, an hour) and from 02:36 to 02:51 UTC on 16 October (major), so several majors (0). Rate limits are described per organisation in requests and tokens per minute and per day, with concurrency caps and headers, but no figures are published and the beta allocation isn't stated (5 of 15). 429 and 503 responses carry `Retry-After`, daily limits add `x-should-retry: false`, and the errors page says which codes to retry with backoff and jitter. Chat completions write nothing (15). No SLA found (0). The API is a public beta per the launch post and the beta pricing page (0). 40.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-reliability):

Hosted APIs, MCP servers, models and platforms.

- 20, a public status page with component history (Statuspage, Instatus, BetterStack or the vendor's own).
- 0 to 30, the incident record for the last 90 days on that page. 30 for a clean record or trivial incidents only, 20 for minor incidents only, 10 for one major outage (an hour or more of a core API down, or errors across the board), 0 for several. 5 when there's no history we could read, and the note says so.
- 15, rate limits documented with numbers.
- 15, documented 429 or overload handling (Retry-After, backoff guidance), and idempotency keys or safe-retry guidance where writes are involved.
- 10, an SLA published for any paid tier.
- 10, the surface agents use is generally available, not beta or preview.

Local packages, SDKs, frameworks and stdio MCP servers.

- 20, installs from an official package with supported runtimes stated.
- 25, a public CI and test suite, passing on the default branch.
- 0 to 25, open crash or regression issues relative to activity (25 for few and handled, 0 for many, old and unanswered).
- 15, semver discipline and breaking changes called out in a changelog.
- 15, version 1.0 or later, or declared stable.

Protocols are read from their reference implementations, the public facilitators or servers, spec stability and test vectors.

## 2. Payments & pricing, 20 out of 100, up to 10 more on the total

Why it scored 20: Graded on the Reflection API. No x402, MPP or L402 (0). The public beta is free with a daily allocation, but no allocation figures or per-token prices are published, and the beta pricing page still shows template placeholders, so plan-level only (10 of 20). Free to start, but the docs say some organisations must verify a credit card before keys work (10 of 20). A person signs up in the browser at platform.reflection.ai to get a key (0). The weights are free under Apache-2.0 for anyone serving them, and hosted copies through other providers are not graded. 20.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-payments):

The published rubric, also on the [x402 page](https://www.anchorterminal.com/x402/).

- 40, a machine payment protocol (x402, MPP or L402) on the tool's own endpoints. 10 to 30 when it covers only some endpoints or only goes through a third party, and the note says which.
- 20, per-call or per-unit pricing published without a login. 10 for public plan-only pricing, 0 for "contact sales" or prices behind a login.
- 20, a free tier or trial that doesn't need a card.
- 20, autonomous onboarding, meaning an agent can get access without a person signing up in a browser (keyless use, x402, a programmatic key API).

Payment platforms and agent wallets rarely charge for their own API over a machine protocol, so the first line has steps for them, and the highest one that applies counts. 40 when x402, MPP or L402 runs on all their own endpoints, 30 when it runs on part of their own API, 25 when their merchants can accept one, 20 for running a facilitator, 15 for paying as a buyer, and 0 when the only protocol is their own. Merchant acceptance sits above a facilitator because the platform's own customers can charge agents through it, while a facilitator settles for sellers who wire up the protocol themselves. The counter-argument (a facilitator does more for the protocol as a whole) has a point. Each note says which step applied.

Open-source software you run yourself is scored on its hosted or paid option if it has one. A free, self-hosted package with nothing to buy gets 20, 20 and 20 for the last three lines, and 0 to 40 for the first only if it ships a payment protocol.

## 3. Security & auth, 46 out of 100, up to 9.5 more on the total

Why it scored 46: Model reading. API keys per project, revocable, shown once and disabled when their creator leaves, with no scopes, expiry or rotation API found (20 of 30). Training is contradictory. The API safety page says content isn't used for training by default, with an opt-in, but the commercial terms of 5 October grant Reflection a licence to train on content sent to Discounted Services, which they define to include services at no charge, and the API is currently free. A `reflection-data-retention` header reports `contributor` when training applies (10 of 20). Content kept 30 days for safety review. Zero data retention isn't available yet, though the header reports the retention mode on each response (5 of 15). `x-request-id` on every response and the retention header, with usage dashboards behind the login and not read (5 of 15). No security.txt on reflection.ai, developers.reflection.ai or api.reflection.ai, no bug bounty of Reflection's own and no SOC 2 or ISO 27001 found. A vulnerability address on the model card, a published safety report with indirect prompt-injection results, and Reflection co-sponsors Gray Swan's public prompt-injection challenge (6 of 20). 46.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-security):

- 0 to 30, the credential model. 30 for OAuth 2.1 with scopes, or scoped and revocable keys with rotation. 20 for plain revocable API keys. 10 for one all-powerful key. 10 off when a secret can travel in a URL query string as a documented option.
- 0 to 20, read-only or least-privilege modes, and confirmation or approval for destructive actions.
- 0 to 15, prompt-injection posture where the tool returns untrusted content (documented mitigations or guidance). A tool that returns no untrusted content gets 10.
- 0 to 15, audit logs or per-call visibility for the operator.
- 0 to 20, a security programme. security.txt or a disclosure policy, a bug bounty, SOC 2 or ISO 27001, advisories handled in public.

Models are read for retention, whether API data trains models (and whether that's off by default), zero-retention options and certifications. Frameworks for telemetry defaults, approval hooks, guardrails and sandboxing.

## 4. Agent ergonomics, 67 out of 100, up to 5.4 more on the total

Why it scored 67: Model reading (tool use, structured output, caching, context, batch, SDKs and defaults, errors). `tool_choice` of auto, none, required or a named function, `parallel_tool_calls`, strict tools and up to 128 tools. The tool-calling loop needs `reasoning_content` sent back on every assistant turn, and the docs say the model sometimes returns a tool-call step without it, after which every request in the session fails (12 of 20). `json_object` and strict `json_schema` (15). Cached prompt tokens are reported in `usage`, but caching can't be steered, the `prompt_cache_*` options are ignored and no cached price exists in the beta (5 of 10). 1,048,576-token context and 131,072 output tokens (15). No batch endpoint (0 of 10). No Reflection SDK, the OpenAI Python and TypeScript SDKs are documented, and defaults cost an agent turns, with reasoning always on at `max`, `stop` and `seed` ignored and a default output budget not reduced to fit the prompt (5 of 10). Machine-readable `error.code` values with `param`, `x-should-retry`, `Retry-After` and `x-request-id` (15). 67.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-ergonomics):

- 0 to 25, context cost. For MCP, the number and size of the tool definitions (25 for ten or fewer compact tools, 15 for 11 to 30, 5 for more than 30, plus up to 10 back for toolsets, dynamic loading or read-only subsets). For APIs, whether responses can be sized (field selection, limits, summaries).
- 20, pagination, filtering and output-size controls.
- 20, actionable, documented error responses, codes and messages an agent can recover from.
- 20, idempotency or safe retries, and for MCP the `readOnlyHint` and `destructiveHint` annotations.
- 15, sensible defaults, few required parameters, and official SDKs in at least two languages.

Models are read for tool use, structured output, prompt caching, context length, batch and SDKs. Frameworks for how much code and how many defaults a tool-calling agent with MCP needs.

## 5. Maintenance & community, 50 out of 100, up to 4.4 more on the total

Why it scored 50: Model reading for a model released the day before this check. Weights and the public API released on 16 October 2026 (30). One model, with a `shutdown_date` field and a lifecycle section in the docs but no notice period for retirements, and no history of releases or retirements yet (4 of 20). No changelog. Support goes through the platform and Discord, and the model card names reports@reflection.ai. Hugging Face discussions not sampled (6 of 25). No first-party SDK, the docs rely on the OpenAI SDKs (5 of 15). Day-one support in vLLM, SGLang, Transformers, TokenSpeed and Dynamo with Reflection's recipes, and a Beam Verifier repository, but fine-tuning recipes are 'coming soon' (5 of 10). 50.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-maintenance):

- 0 to 30, time since the last release, or the last published model or API change for a closed service. 30 within 30 days, 20 within 90, 10 within 180, 0 older.
- 20, at least three releases or dated changelog entries in the last 90 days.
- 0 to 25, responsiveness. Issues and pull requests answered on GitHub (the open issues and how recent the replies are). For closed services, a public changelog and a support or community channel that answers, 0 to 15.
- 15, presence in the official MCP registry under a verified namespace (MCP servers), or current official SDKs (APIs and models).
- 10, package health, current dependencies and CI.

Models are read for deprecation notice periods and model churn rather than release counts.

## 6. Transparency & trust, 68 out of 100, up to 2.8 more on the total

Made of editorial 59, provenance 76.

Why it scored 68: Weights under Apache-2.0, an OSI licence, with no restriction on distillation or training on outputs. The hosted API is closed under clear commercial terms, and training code and data aren't released (25 of 30). Data handling statements disagree. Flagged content is kept up to 12 months in the terms, up to two years on the API safety page and 'one year' later on that page, training is off by default on one page while the terms license training on free use, and the docs reset daily limits at 00:00 UTC while the pricing page says a bracketed midnight Pacific time (10 of 30). A `shutdown_date` field and 30 days' notice before fee increases, but no model deprecation policy or dated notices (6 of 20). A subprocessor list of twelve with purposes and locations, including Together AI in Canada and Voltage Park for inference (18 of 20). 59.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-transparency):

- 0 to 30, source availability and licence clarity. 30 for open source under an OSI licence, 15 for closed with clear terms, 0 for unclear terms.
- 0 to 30, data handling and retention statements that agree with each other (privacy policy, DPA, retention periods, subprocessors).
- 0 to 20, a deprecation policy or notices with dates.
- 0 to 20, telemetry disclosed with an opt-out (local software), or subprocessors and data locations disclosed (hosted).

The other half of Transparency and trust is the provenance score, computed from checked facts (below). The category score is the mean of the two.

Provenance checks not met in full (half of this category, computed from checked facts):

- Domain age: reflection.ai, registered 2017-12-16 (8 years) (11 of 15)
- Changelog: not found (0 of 10)
- security.txt: not found (0 of 10)

## 7. Schema & documentation, 84 out of 100, up to 2.6 more on the total

Why it scored 84: Model reading. OpenAPI 3.0.3 document, version 0.1.0, covering all three documented endpoints with typed schemas and error examples. A second 'Reflection client' base URL, https://api.reflection.ai/v1, is named with no endpoints documented (22 of 25). llms.txt, a Markdown copy of every page and a docs MCP endpoint (10). The guides say which reasoning effort suits which task, which OpenAI parameters are fixed, ignored or unsupported and which endpoints don't exist, so a model knows what not to send (17 of 20). `reasoning_effort` is an enum, strict JSON Schema for tools and `json_schema` output with the unsupported keywords and size limits listed, and unknown parameters return `unsupported_parameter` (15). curl, Python and TypeScript examples on every guide, and an error reference with every status, `error.code` and retry advice (15). The model ID carries its version and the models endpoint has a `shutdown_date` field, but no changelog or release notes were found (5 of 15). 84.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-schema):

APIs and MCP servers.

- 25, a machine-readable contract (a public OpenAPI file or similar; for MCP, typed JSON Schema inputs on every tool).
- 10, llms.txt or Markdown docs served for agents.
- 0 to 20, descriptions that say what a tool is for, when to use it and when not to, read from the tool definitions in the source or the API reference.
- 0 to 15, typed inputs with enums, constraints and required fields, and no free-form JSON blobs.
- 0 to 15, examples and documented error responses.
- 15, versioning and a public changelog.

Models are read from the API reference, the OpenAPI file, llms.txt, the structured-output and tool-use docs and the model cards. Frameworks from docs a model can follow, typed interfaces, examples and the API reference.

## What we couldn't check

What we couldn't read counted as absent. Publishing it on a page a plain HTTP fetch can read (not only in a browser) lets the next check count it.

- unchecked: the platform's usage dashboard, billing tiers and any per-token price, which sit behind the login at platform.reflection.ai
- unchecked: the size of the beta's daily allocation and the per-organisation rate limits, which no public page states
- unchecked: Hugging Face discussions and GitHub issues on reflection-oss repositories, not sampled
- The commercial terms' training licence for content sent to free services and the API safety page's 'not used for training by default' disagree. Recorded as a transparency and security weakness, not a deduction. The founder may want to ask Reflection which governs the free beta
- firstReleased is 16 October 2026, the public release of the weights and API. The models endpoint gives 5 October 2026 as `created`, the date of the announcement, when access was early access only

## Weaknesses

- The API is a public beta with no SLA, and per-organisation rate limits and the daily allocation are not published as numbers
- Reasoning is mandatory and defaults to `max`, and a tool-call turn returned without `reasoning_content` makes every later request in that session fail
- The commercial terms grant Reflection a training licence over content sent to free or discounted services, while its API safety page says content isn't used for training by default
- Flagged content retention is 12 months in the terms but two years on the API safety page, and zero data retention is not yet available
- Three API incidents in its first three days, two rated critical, with no changelog or model deprecation notice period found

## What costs an agent a turn today

The notes we give agents before they call it. Each one is a workaround an agent shouldn't need.

- Send `reasoning_effort` `low` or `medium` for short tasks. The default is `max`, and `none` or `minimal` return `400 unsupported_value`
- Return every assistant message with its `reasoning_content` unchanged. If a turn comes back without it, start a new conversation, since later requests fail with `missing_required_parameter`
- Don't rely on `stop` or `seed`. Both are accepted and ignored, so truncate output in your own code
- Honour `x-should-retry: false` on a 429. It means a daily limit is spent and `Retry-After` counts to 00:00 UTC
- Set `max_completion_tokens` yourself on long prompts. The default budget isn't reduced to fit the prompt and includes reasoning tokens

## What the review panel asked for

- Publish allocation sizes and token prices
- Return an error for ignored parameters
- Publish a changelog
- Publish daily allocation and rate limits
- State a beta end date or notice period
- state which training rule governs the beta
- zero data retention

## When it's done

Send what changed and where it's published as a dispute (https://www.anchorterminal.com/builders/#disputes, or `POST https://www.anchorterminal.com/api/v1/contact` with `"kind": "dispute"`). Disputes are answered in public, and the listing is checked again by the same checklist. Paying for an audit or a listing claim changes nothing here.