# Fix list: pplx-decider-flash From Anchor Terminal's listing at https://www.anchorterminal.com/tools/pplx-decider-flash, the October 2026 research run, assessed 17 October 2026. Grade B, 67.9 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 pplx-decider-flash: 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. Payments & pricing, 25 out of 100, up to 9.4 more on the total Why it scored 25: No x402, MPP or L402 in the guide, the reference, the pricing page or llms.txt (0). The price is published without a login, $0.01 per million input tokens with free output and no per-request fee (20). Usage is prepaid, the billing guide asked for a credit card on 9 October, and no free allowance was found (0). A person signs up in the console. The Stripe Projects route needs a person's Stripe account with a payment method and a $10 minimum purchase, per the 9 October check (5 of 20). The weights are free, and the hosted price is what this line grades. 25. 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. ## 2. Reliability, 66 out of 100, up to 6.8 more on the total Why it scored 66: Read as a hosted model API, the same endpoint as pplx-decider. status.perplexity.com has an API component with uptime history and an incident feed. It has no Decisions component (20). The Decisions API was released in October 2026 and Flash on 16 October, so the record is about two weeks long for the endpoint and a day for the model. The feed's one incident since the API's release, on 9 October 2026, says the API remained operational, so 8 of 30 for a clean record too short to count, the reading given to the other decision APIs in their first days. The guide states 10 requests a second per organisation on every plan, shared with the 27B model, and a token limit with no number (13 of 15). 429 carries `Retry-After`, successful responses carry `x-ratelimit-*` headers, the guide says to retry 5xx with backoff, and a decision has no side effects (15). The API FAQ says Perplexity does not guarantee uptime or recovery time (0). No beta label in the guide, the reference or the OpenAPI description at version 1.0.0, and both official SDKs have a Decisions method (10). 66. 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. ## 3. Security & auth, 68 out of 100, up to 5.6 more on the total Why it scored 68: Read for a model API, on the same account, keys and terms as pplx-decider. Project API keys are shown once, can be revoked, and can be generated and revoked through two documented endpoints, per the 9 October check. No per-key scopes or spending limits were found (22 of 30). The API terms say Perplexity will not use Customer Content to train any generative AI model, and the DPA says personal data will not train its models. The FAQ says API query data is not retained, while the Privacy and Security page says zero data retention is not the default and needs a written agreement (13 of 20). The endpoint returns only probabilities and never fetches a URL, and the card says literal special tokens in user content are neutralised in the published inference code. The hosted docs have no guidance on adversarial input (6 of 15). Responses carry `x-request-id` and token usage, and admins see usage per key in the console. No audit log was found (10 of 15). security.txt on www.perplexity.ai is valid until 31 December 2027, the security page names a public vulnerability disclosure programme, a private Bugcrowd bounty, annual penetration tests and SOC 2 Type II, and the trust centre is script-drawn and was not read (17 of 20). 68. 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, 81 out of 100, up to 3.1 more on the total Why it scored 81: Read as an API an agent calls for a decision. Answers are a few numbers per question and input may run to just under 256K tokens. The API bills the state once for all the questions in a request, but Perplexity advises one question per request for Flash, so following that advice bills the state again for every question. No batch route or caching was found (18 of 25). The caller sets the whole output shape (18 of 20). The error table gives a cause and an action for each status. 404 and 405 have empty bodies, 504 can be HTML, and an oversized image waits about a minute before a 504 (14 of 20). A decision has no side effects and 429 names its wait. The same question can score differently depending on what it is grouped with, which the guide documents (17 of 20). Three required fields. The official Python SDK 0.44.0 and TypeScript SDK 0.39.0 add `client.decisions.create`, though their examples name the 27B v1.1 model (14 of 15). 81. 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, 67 out of 100, up to 2.9 more on the total Why it scored 67: Read for a hosted model with open weights. Flash v1.2 was committed on 16 October 2026 and announced in the changelog under October 2026 (30). The guide states that a retired version's name routes to its replacement, announced in the changelog, with no notice period, and on 16 October the 27B v1.1 and v1 names were moved to v1.2 the day it shipped (3 of 12). Flash has one version, too short to judge its own churn, and the `pplx-decider-flash-latest` alias can move (4 of 8). The changelog has several entries a month, and support runs through a community forum, a GitHub discussion repository and email. Replies were not sampled (10 of 15). The official Python SDK 0.44.0 and TypeScript SDK 0.39.0 carry Decisions methods from 15 October (14 of 15). The repository pins its requirements (torch 2.14.0, transformers 5.17.0) and has an Apache-2.0 `pyproject.toml`, with no lock file or CI, and its card still calls the repository private (6 of 10). 67. 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, 71 out of 100, up to 2.5 more on the total Made of editorial 60, provenance 82. Why it scored 71: The weights and the `pplx_decider` inference package are public under Apache-2.0, with a `NOTICE` naming Qwen/Qwen3.5-4B and its revision. The card says how the model was made, by distillation from the 27B models and weight averaging, but no training code, config or data manifest is published, and the card and release manifest still say private. The hosted service is closed (26 of 30). The API terms and DPA agree that customer content does not train models, and the DPA deletes personal data within 30 days of the service ending. The FAQ's statement that API query data is not retained contradicts the Privacy and Security page, which says zero data retention is not the default (16 of 30). Retirements are announced in the changelog and a retired name routes to its replacement, with no notice period (8 of 20). The FAQ says compute is on AWS in North America. The DPA points to a subprocessor list on the script-drawn trust centre, which was not read, and promises ten business days' notice of new subprocessors other than model providers (10 of 20). 60. 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: perplexity.ai, no registry record we could read (0 of 15) - Privacy policy: published, but our reader couldn't read it (7 of 10) ## 7. Schema & documentation, 89 out of 100, up to 1.8 more on the total Why it scored 89: The same documentation as pplx-decider, which covers both models. An OpenAPI 3.1.0 description of `POST /v1/decisions` in docs.perplexity.ai/openapi.json, listed in llms.txt and embedded in the reference page (25). llms.txt with Markdown twins of the guide, the reference and five cookbook pages (10). The guide says what a decision model is and what it does not do, and when to pick Flash over the 27B model, with a per-model note on when grouping questions hurts. No page lists languages (18 of 20). Questions are a union discriminated on `type`, with 1 to 128 questions, 1 to 255 options, 1 to 10 levels and unknown fields rejected. `model` is a free string, and `state` is free-form by design (13 of 15). Examples in Python, TypeScript and curl and a status table for 400, 401, 404, 405, 413, 429 and 5xx (14 of 15). A versioned name and an alias, a `/v1` path and a changelog dated by month. The guide says a retired name routes to its replacement, and the response echoes the name sent rather than the version that answered (9 of 15). 89. 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 announcement the founder forwarded ('top 4B model on HF Decision Index', 1 cent per million tokens) was not found on a first-hand page we could read. The size, price and availability were confirmed from the changelog, the guide and Hugging Face - The 'top 4B model' claim. On Decision Index 0.3.2 (16 October 2026) Flash has the highest Full score, 54.81, of the entrants of about 4B by the index's served-parameter counts, ahead of vLLM-SR d3-mini at 52.61. It also leads them on the public-only part (60.06) and on the vision board (66.02). The claim holds on the index as read on 17 October 2026, so no deduction - The 'HF Decision Index' is the Space multimodalart/jev-decision-index under a personal account, which its README calls unofficial and community-maintained. 80 per cent of its Full score comes from private tests nobody outside can inspect. We have not run the model - The index ran Flash with one question per request, the submitter's chosen mode. The card's 58.88 is in joint mode and its 60.05 in one-question mode - Whether Flash's training data includes Decision Index training splits, as the 27B v1.2 card says for that model, is not stated on the Flash card. It is distilled on the 27B models' training requests - The card and `release-manifest.json` call the repository private, and the index's entry says the same, while Hugging Face reports it public and not gated on 17 October 2026 - The hosted API accepts just under 256K input tokens. The card says training used sequences up to about 32k tokens and longer inputs were only spot-checked - The terms read are the Perplexity API Terms of Service of 23 January 2026, which do not name the Decisions API. Whether product-specific terms apply was not established - unchecked: trust.perplexity.ai is script-drawn, so the SOC 2 Type II report and the subprocessor list were not read - unchecked: the billing guide, Stripe Projects page and key-management page were not re-read on 17 October. The 9 October check of pplx-decider is relied on for them - unchecked: pypi.org answered with a client challenge, so the SDK release dates come from the repositories' changelogs and tags and from npm's registry - Whether the hosted service runs the published `pplx_decider` code was not established. The hosted API caps a request at 128 questions and a `score` at 10 levels, and the code has no question cap and allows 255 levels ## Weaknesses - Released on 16 October 2026, with no service record of its own and no Decisions component on the status page - Grouping questions costs up to 20.1 points on one benchmark by Perplexity's figures, so one question per request is the advised default - On the Decision Index's Full score it trails the 27B v1.2 by 10.74 points, 54.81 against 65.55 - The `pplx-decider-flash-latest` alias can move to a newer model, and a retired name routes to its replacement with no notice period - The FAQ and the Privacy and Security page disagree on whether API data is retained by default ## 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 `Authorization: Bearer` with `model` set to `pplx-decider-flash-v1.2`. A key in `x-api-key` returns 401. - Send one question per request by default. Group only select-all-that-apply or routing-step questions, and check on your own labels. - Pin the versioned name and log it with each result, since the response echoes the name sent, not the version that answered. - Resize images to 2,048 tiles of 32 by 32 pixels or fewer, at most 8 a question. Larger images end in a 504 after about a minute. - Compare it with `pplx-decider-v1.2` on a labelled sample at your thresholds before choosing it. ## What the review panel asked for - dated retirement notice period - answering version in responses - Per-key spending limits - Machine-payable access - JSON bodies on 404, 405 and 504 - Return the serving version in responses - Publish the token limit figure - Add a Decisions status component ## 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.