Perplexity news & updates

Perplexity expands its API model list and removes older documentation entries

Perplexity’s Agent API now documents GPT-6 Astra and a Perplexity-hosted DeepSeek V4 Pro 0813 endpoint. The DeepSeek endpoint is also listed for its Orchestrator API. The same notice removes 16 older Anthropic, Google and xAI IDs from the documentation, pricing calculator and examples, and recommends current replacements. It does not announce their shutdown or establish that existing API calls fail.

Why it matters: The catalogue changes affect integration choices, but documentation removal should not be mistaken for a confirmed retirement.

Perplexity’s official RSS dates the notice to 17 October 2026 at 12:16 BST. The first callable moment is not separately specified.

Perplexity publishes Decider 1.2 and a smaller Flash decision model

Perplexity has published Apache-2.0 checkpoints for the 27-billion-parameter Decider 1.2 and a new four-billion-parameter Decider Flash 1.2, which accept text and images and return decision probabilities rather than conversational prose. Their joint-question format can connect related decisions, but the model cards warn that grouping independent candidates can also hurt results. The API lists input-token prices of US$0.02 per million for Decider 1.2 and US$0.01 for Flash, with free output tokens and older Decider identifiers redirected to 1.2. Company-reported evaluations include training on official benchmark training splits with test-set decontamination, so they should not be read as independent, wholly out-of-domain validation.

Why it matters: The smaller checkpoint adds a new deployment option for bounded decisions, while grouping effects, benchmark methodology and model-ID redirects matter for existing integrations.

Perplexity announced the API/model update on 16 October 2026; the Hugging Face repositories were created on 15 October and updated on 16 October. The US$0.02 main-model input price was already documented for version 1.1; this is not a newly claimed price cut.

Perplexity broadens Agent API inputs and changes search defaults

Perplexity’s 15 October API ledger adds native PDF, DOC, DOCX, TXT and RTF inputs and an image-search tool that returns typed results with domain, format and safe-search controls. Low-, medium- and high-effort Agent presets now use Fast Search by default, cutting the listed web-search tool fee from $2.50 to $1 per 1,000 calls; xhigh remains on standard search, and callers can opt back in with `search_type: "web"`. Document requests cannot use browser tools and are unavailable under Zero Data Retention, while image search costs $2.50 per 1,000 successful invocations.

Why it matters: Agents can now analyse common office files and retrieve images natively, while the default-search change materially affects latency, cost and retrieval behaviour for existing integrations.

The official changelog metadata was updated at 18:31:11 UTC on 15 October, after this tracker’s prior check. The detailed tool guides were published earlier, so that timestamp establishes the consolidated availability notice rather than the precise activation moment for every feature.

Perplexity moves Decisions API to v1.1 and halves input price

Perplexity lists `pplx-decider-v1.1-27b` as replacing v1 in the Decisions API and cuts input pricing from $0.04 to $0.02 per million tokens. Existing `pplx-decider-v1-27b` requests continue to work and receive the same lower rate, while output remains free. The endpoint returns typed probabilities for yes-or-no, multiple-choice or scored decisions; it does not generate an explanatory rationale.

Why it matters: The lower price and in-place compatibility reduce adoption friction for applications that need bounded, machine-readable choices rather than open-ended text generation.

Recorded in Perplexity’s official changelog update at 18:31:11 UTC on 15 October. Perplexity supplies no separate benchmark or migration note quantifying how v1.1 differs from v1, so the entry does not infer a quality improvement.

Perplexity open-sources its SPACE agent-sandbox platform

Perplexity announced that SPACE — the Sandboxed Platform for Agentic Code Execution used behind its Computer product — is now open source. The Apache 2.0 repository uses Firecracker microVMs and exposes APIs for creating isolated Linux environments, running commands, streaming output and managing their lifecycle, with a local Kind setup, a reference AWS EKS deployment and an E2B-compatible layer. The repository is usable source code, but its security and production-hardening claims have not been independently audited here.

Why it matters: Teams building coding and research agents can inspect and self-host the isolation layer instead of treating remote execution as an opaque service.

Perplexity’s article index dates the announcement 15 October but exposes no publication time; repository history predates the announcement, so this entry concerns the public open-source announcement rather than claiming the code was first created that day.

Perplexity releases multimodal retrieval models

Perplexity released two pplx-embed-v2-late models for searching text, images and visual documents. Their shared embedding space lets the smaller model search an index created with the larger one. The weights are available on Hugging Face under an MIT license; these are retrieval models rather than new chat models.

Why it matters: The release offers document and search tools a way to balance retrieval quality with query speed.

Perplexity Agent API adds reusable, versioned profiles

Perplexity’s Agent API now lets teams save an agent’s model, instructions, tools, Skills and managed connectors under one versioned Profile ID. The saved configuration can be reused across applications and workflows. The official announcement describes an API Portal feature and does not establish that the same controls are available in the consumer app.

Why it matters: Reusable profiles help teams apply a consistent agent setup across several applications.