Perplexity September 2026 Update Adds Effort Mode and GPT-6 Astra

What Perplexity shipped in September

In September 2026, Perplexity spent less effort on its answer engine and more on Computer, its agent product. The update that matters most is dated September 21. It added an effort-level slider to Computer, GPT-6 Astra on the High setting, a Skills Marketplace, Hybrid compute that mixes cloud and local models on Mac, and Side Chat, which lets users ask questions while a task keeps running.1 The tracker Releasebot recorded the same release and gave sample prompts for each effort level. A Light-effort task might be summarizing files to pull out the decisions a user needs to make. Heavier tasks might involve researching options, building a comparison table and making a recommendation.2

The four effort levels are Light, Standard, High and Ultra. Users can also skip the presets and pick the model and reasoning depth themselves.10 Perplexity says the controls match both the model and the amount of reasoning to the job.9

More releases followed. On September 1, Hybrid Compute on Mac started Computer tasks in the cloud and kept sensitive files on the device.9 One release tracker reported that the local model is PPLX Qwen 3.8 27B. It runs on Apple silicon behind an on-device classifier that flags personal data, and local work does not use cloud credits.14 Portable Computer for Windows arrived in mid-September. HP agreed to pre-load the Perplexity Windows app on its PCs. Portable Computer also came to AMD Ryzen AI Max machines, so local and recurring tasks can run without spending credits.9 Late in the month, Perplexity added Automations, which run Computer tasks on a schedule or when an event triggers them. Each run builds on the previous one.5

The developer side: Sonar retires, the Agent API takes over

For developers, the biggest change was the end of Sonar. Perplexity's docs said the Sonar chat completions endpoint would be supported only until September 27, 2026. New projects were pointed to the Agent API instead.3 The routing service LLM Gateway described how the move played out for its customers. On September 25, plain sonar moved to the Agent API with the same requests and a lower price. On September 27, sonar-pro and sonar-reasoning-pro stopped working because the Agent API rejects them and has no direct replacement.7 An analysis by NeuralTrust confirmed the September 27 cutoff. It added that synchronous and streaming requests are converted into Agent API calls.16

Pricing changed alongside the shutdown. Perplexity's API changelog says the fast, low, medium and high presets now use Fast Search. That cuts web search from $2.50 to $1.00 per 1,000 calls and makes search about 800 milliseconds faster. The xhigh preset still uses standard search.4 NeuralTrust's comparison table, retrieved October 5, still lists Agent API search at $2.50 per 1,000.16 The most likely explanation is that NeuralTrust used the standard rate while the presets had already moved to the cheaper tier. Developers should check which preset they actually use before trusting either number.

Perplexity also added an Orchestrator API. It serves open-weight models that Perplexity hosts, works with both OpenAI Chat Completions and Anthropic Messages formats, and routes traffic around unhealthy deployments automatically.4 At the end of September, the Agent API gained reusable agents: versioned Profiles, Skills and managed connectors that teams can share across applications.5

The Anthropic angle: a key supplier that also competes

Anthropic's own September was crowded. Anthropic released Claude Fable 5.1 and Mythos 5.1 on September 1. Anthropic describes them as the same model with different levels of safeguards, and Mythos 5.1 is limited to selected programs.14 Claude Opus 5.5 followed on September 22 and Sonnet 5.5 on September 28.12 One comparison site reported that Opus 5.5 took first place on the Artificial Analysis Intelligence Index.13 Anthropic itself cautioned at launch that benchmark gaps have become a weaker guide to real-world differences.12

Perplexity's ties to Anthropic run through almost every layer of the product. In July, its changelog brought Claude Opus 5 to Search and Computer.2 Earlier releases used Claude Fable 5 to coordinate Computer tasks and added a faster mode for Opus 4.8.2 By early October, the Agent API listed Opus 5.5 and Sonnet 5.5 at the same prices Anthropic charges: $4/$20 and $2/$10 per million tokens.16 On the consumer side, reports from September put Claude Opus 5 behind the $200 Max plan, with Sonnet 5 available on Pro.17 One comparison noted that Perplexity's published model list was dated September 4 and did not yet include Anthropic's 5.5 models.12

The headline model of Perplexity's September release came from OpenAI, not Anthropic. GPT-6 Astra went on Computer's High setting.1 An October roundup said GPT-6.1 Sol now powers the Light setting.10 Earlier in the year, OpenAI models were also the defaults for coordinating Computer tasks: first GPT-5.5, then GPT-5.6 Terra for sub-agents.26

My reading is that Perplexity is avoiding dependence on any single lab. Effort Mode makes that choice deliberate. Light, Standard, High and Ultra are labels Perplexity owns, and it can quietly change which lab's model sits behind each one. Anthropic's most capable models are still on the menu, but they are now options within Perplexity's own framework rather than the main attraction.

The competitive pressure goes both ways. Comparison coverage describes Claude as increasingly agent-driven. Cowork works on a user's files, and Claude in Chrome operates the browser.11 That is the same territory Perplexity Computer and its Comet browser are going after. One clear difference: a Perplexity subscriber can use Claude inside Perplexity, but Claude's own apps run only Anthropic models.16

The research Perplexity published

Perplexity's research team published several papers in September that help explain the product strategy:

• "Learning from real-world experience" (September 21): describes further training a Computer model to copy successful behavior, need fewer user corrections and avoid tool errors.5
• "Escaping SPACE: Part I" (September 23): a red-team study of whether AI agents can break out of virtual machine isolation and network limits.5
• Photon (September 24): a retrieval and ranking engine built from scratch, using specialized data formats and keeping index building separate from query serving.5
• Safer agents (September 29): describes layered safeguards, independent monitoring and human oversight across Perplexity's cloud, browser, local and endpoint agents.5

These papers line up with the product launches. Hybrid and Portable Computer move agents onto users' own machines, so containing those agents becomes a real engineering problem. Photon reflects a company that still treats search as its core advantage, even as it rents the reasoning models from other labs. The real-world training work points to a model trained on Perplexity's own Computer data. That could reduce how much a given frontier model matters, at least on routine tasks.

In early October, Perplexity released an open-weight 27B model behind a new Decisions API under the Apache 2.0 license, followed by an updated v1.1 version.10 The API changelog says v1.1 replaces v1 and lowers input pricing from $0.04 to $0.02 per million tokens.4

Where the coverage disagrees

The dates don't always line up. Perplexity's news page dates the effort controls to September 17, but the main changelog entry is dated September 21.91 Automations appeared on Perplexity's blog on September 29, while an October roundup dates them to October 5.510 These look like staggered rollouts rather than conflicting facts, but the pattern matters. One tracker noted that Perplexity's own pages sometimes disagree with each other and that its plans change often.3

The bigger gap is between product momentum and market traction. One comparison site estimated that Perplexity's share of chatbot traffic fell from 5.3% in January to 1.9% in September, while Claude's rose from 8.5% to 10.6%.13 That is a single estimate and should be treated with caution. Still, it fits the bet Perplexity appears to be making with these launches. As frontier models become more alike and Anthropic and OpenAI build their own agents, a search-first product that relies on other labs' models has to compete on coordination, running tasks on local hardware, and giving users controls they can trust.

September showed Perplexity becoming an agent platform that uses other labs' models, with its own small models and search technology underneath. The open question is whether users will value that layer more than the models running inside it.