All issues
Discovery Digest · October 12, 2026
Issue 20. The week AI went interactive and search started paying publishers
TL;DR
OpenAI rolled out GPT-6 Intelligent UI to all ChatGPT users on 7 and 8 October, replacing text-only answers with interactive charts, calculators, and functional forms across every plan tier. Anthropic cut Claude Haiku 5.5 pricing to 75 per cent cheaper per task than its predecessor and added effort control to the Haiku range for the first time, while Perplexity reduced its Agent API web search cost by 60 per cent and released pplx-decider-v1, an open-weight 27B model that returns structured decisions rather than text. Against that backdrop of falling costs and richer interfaces, Google confirmed the most significant development in search economics this cycle: an invitation-only pilot paying around 100 publishers monthly for content used in AI Overviews and AI Mode, with payouts varying from under $1,000 to more than $1 million per year and no published formula for how they are calculated.
01 · OpenAI
OpenAI rolls out GPT-6 Intelligent UI to all ChatGPT tiers, replacing text with interactive charts, calculators, and forms
- What
- OpenAI began rolling out GPT-6 with Intelligent UI on 7 October to Plus, Pro, Business, and Enterprise users (all receiving GPT-6 Sol), followed by Free and Go users on 8 October (receiving GPT-6 Luna). Intelligent UI enables ChatGPT responses to include interactive and visual components selected automatically by the model: side-by-side comparisons, charts, calculators, bill splitters, quizzes, games, forms, and maps appear when the model determines a visual format serves the answer better than text. The model also begins responding while still reasoning, which OpenAI says reduces perceived wait times by 44 per cent.
- When
- Rollout to paid tiers began Tuesday 7 October 2026, extending to Free and Go users on Wednesday 8 October. Reported by TechRepublic and The Decoder.
- How it shifts discovery
- The commercial implication is not the visual design. It is that ChatGPT has moved from answering questions in text to completing tasks in interface. A user asking about mortgage repayments no longer gets an explanation: they get a working calculator. A user researching a trip gets a map alongside the itinerary. This compresses the gap between the conversational layer and the application layer: tasks that previously required a user to leave ChatGPT for a dedicated tool now complete inside it. The parallel with Google's Generative UI expansion into standard AI Overviews from August 2026 is exact: both platforms are now building functional equivalents of brand-owned tools inside their own interfaces, without routing the user anywhere. For any organisation whose conversion assets include a calculator, configurator, comparison table, or quiz, the window to audit which of those tasks is now completable inside ChatGPT is this month.
- Questions to ask
- Intelligent UI builds calculators, comparison tables, interactive charts, and forms inside ChatGPT automatically based on query context. For the highest-traffic queries in our category, have we tested what GPT-6 Intelligent UI now produces in response, and does any output compete with or displace our own interactive tools or conversion assets?
- GPT-6 Sol rolls out to paid tiers and GPT-6 Luna to free tiers, with the model adapting output format automatically. For organisations using ChatGPT Business or Enterprise, has a named owner assessed what interactive outputs the model now generates for employee queries, and are there categories of query where an auto-generated chart or calculator introduces accuracy or compliance risk?
- The shift from text output to interactive output changes what it means to appear in a ChatGPT answer: a citation in a text response is different from being the data source behind a calculator a user interacts with directly. Has a named owner defined what brand presence looks like inside Intelligent UI outputs, and is there a monitoring process for how our brand and category appear in interactive responses?
- Sources
02 · Anthropic
Anthropic launches Claude Haiku 5.5 at 75% lower running cost with effort control, targeting high-volume AI workloads
- What
- Anthropic released Claude Haiku 5.5 on 7 October, priced at $0.10 per million input tokens and $0.50 per million output tokens, which Anthropic says costs around 75 per cent less to run per task than Haiku 4.5. The model is Anthropic's fastest to date. For the first time in the Haiku range, developers can adjust the effort the model applies per task, trading output quality against speed and cost without switching to a different model. Anthropic simultaneously halved Sonnet 5.5 cache-read pricing, reducing the cost of repeated reads in context-heavy workflows.
- When
- Released on anthropic.com, Tuesday 7 October 2026. Reported by VentureBeat and 9to5Mac.
- How it shifts discovery
- Haiku 5.5 is positioned for high-volume work: document classification, summarisation, database queries, and sub-agent tasks in larger agentic pipelines. At $0.10 per million input tokens it sits at the same price point as GPT-6 Luna, making the comparison with OpenAI's high-volume model direct and deliberate. The effort control feature is the architectural addition that matters most for production use: it allows cost-to-quality tuning at the task level rather than at the model selection level, which simplifies multi-tier model routing in pipelines that currently direct different task types to different models. For any team running Haiku 4.5 at scale, the migration case is immediate. The Sonnet 5.5 cache-read cut adds a further saving for teams running Sonnet in long-context or repeated-read workflows: that reduction is automatic and requires no code change.
- Questions to ask
- Haiku 5.5 costs approximately 75 per cent less per task than Haiku 4.5 and is Anthropic's fastest model. For every production workflow currently on Haiku 4.5 or on a more expensive model handling high-volume, lower-complexity tasks, have we benchmarked Haiku 5.5 against those specific outputs and calculated the revised cost per task before the next billing cycle?
- Effort control lets developers set compute level per task, enabling cost-to-quality tuning at the routing layer rather than requiring a separate model tier. For teams managing multi-model agentic pipelines, has a named owner assessed whether effort control in Haiku 5.5 simplifies the current routing architecture, and at what quality threshold Haiku 5.5 handles tasks currently delegated upward to Sonnet 5.5?
- Anthropic halved Sonnet 5.5 cache-read pricing at the same time as the Haiku 5.5 launch. For workflows using Sonnet 5.5 with large context windows and repeated reads, have we recalculated the revised cache-read cost and confirmed whether the saving changes the economics of those deployments at current scale?
- Sources
03 · Google Search
Google pays around 100 publishers for AI Overview contributions via Search Console, with payouts from under $1,000 to more than $1 million annually and no published formula
- What
- Google is running an invitation-only pilot that pays publishers a monthly amount when their content contributes to answers in AI Mode, AI Overviews, and Gemini. Around 100 publishers are participating. Monthly earnings appear in a new panel in Search Console. Payouts vary sharply: one participant is on pace to earn more than one million dollars per year, while others have earned less than $1,000 across several months. Google has not disclosed the eligibility criteria, the attribution methodology, or what qualifies a piece of content as a contributing source.
- When
- Pilot ongoing; new payout data reported by The Information and covered by Search Engine Journal and Media Copilot in early October 2026.
- How it shifts discovery
- The pilot establishes two things simultaneously: that Google is willing to pay for content that feeds its AI search products, and that the payment relationship is entirely on Google's terms. No participating publisher knows what separates a million-dollar contribution from a thousand-dollar one under the current methodology. For large publishers, the numbers are potentially material: at more than $1 million annually for one participant in a pilot of 100 sites, the programme has significant revenue implications if it scales beyond invitation. The bigger structural signal is this: after two years of AI search absorbing content without payment, Google has created the first formal mechanism to share AI search revenue with content creators. The formula, the eligibility criteria, and the timeline for a broader rollout are the three unknowns that define whether this becomes consequential for the industry or remains a goodwill gesture toward a handful of partners.
- Questions to ask
- Google's AI contribution pilot pays publishers for content used in AI Mode and AI Overviews, but the attribution methodology is unpublished and payouts vary by more than 1,000 times across participants. For publishers currently in or invited to join the pilot, what is the plan for monitoring whether the payout reflects our content's actual contribution to AI answers, and who owns that relationship with Google?
- One publisher is on pace for more than $1 million per year; others have earned less than $1,000. For large media organisations not yet in the pilot, does the scale of payouts change the case for seeking a direct AI content licensing conversation with Google, or for publicly documenting content's role in AI search answers to strengthen that negotiation?
- The pilot is invitation-only with no published eligibility criteria. Does the absence of a transparent formula change how we prioritise content investment for AI search, and is there a case for structuring new content to optimise AI Overview contribution before a broader version of the programme launches?
- Sources
04 · Perplexity
Perplexity open-sources pplx-decider-v1, a 27B decision model returning probabilities rather than text for agentic AI search pipelines
- What
- Perplexity released pplx-decider-v1 on 1 October, a 27B parameter model fine-tuned from Qwen3.8-27B and published under Apache 2.0 on Hugging Face. The model is built for decision-making tasks where a structured output (a probability score, yes/no, multiple-choice selection, or ordered ranking) is more useful than natural-language text. It processes inputs up to 262,144 tokens. Perplexity released an updated v1.1 with multimodal input support on 6 October and reports 85.7 per cent accuracy on its own 7,210-row evaluation panel.
- When
- pplx-decider-v1 released 1 October 2026; v1.1 with multimodal support released 6 October 2026. Reported by AI Weekly.
- How it shifts discovery
- The Decider model addresses a specific gap in agentic search pipelines: the step between retrieving information and acting on it. Most AI search tools return ranked text that an agent must interpret into a decision. pplx-decider sits differently in the stack: it consumes retrieved context and outputs a probability score or structured choice directly, removing the interpretation layer between search output and downstream action. For teams building AI search workflows where retrieval feeds a recommendation engine, a classification task, or a routing decision, that removal reduces both error rate and latency at the step where agentic pipelines most commonly fail. The Apache 2.0 licence and open weights are the structural advantage: the decision logic is portable, runs on private infrastructure, and is auditable in a way that a closed-API call is not.
- Questions to ask
- pplx-decider returns probabilities and structured choices rather than text, designed to sit between a retrieval step and a downstream decision. For any agentic workflow where search results currently feed a classification, routing, or recommendation task, have we assessed whether structured decision output from pplx-decider reduces the error rate or cost compared with prompting a general-purpose model for the same interpretation?
- The model is Apache 2.0 with public weights, meaning the decision logic is portable and runs on private infrastructure. For organisations with data privacy or regulatory requirements that prevent routing search outputs through third-party model APIs, does pplx-decider's open licence change the feasibility of integrating structured decision intelligence into internal AI search pipelines?
- Perplexity released a v1.1 update with multimodal input on 6 October. For teams building decision workflows that incorporate image or document inputs alongside text, does multimodal support in pplx-decider change the scope of tasks we could route through a structured decision model rather than a general-purpose one?
- Sources
05 · Perplexity
Perplexity cuts Agent API web search cost by 60% and adds Finance Search for live public company data
- What
- Perplexity cut the price of its web_search tool in the Agent API from $2.50 to $1.00 per 1,000 invocations, a 60 per cent reduction. The Agent API fast preset now defaults to Fast Search, the Photon-powered retrieval layer returning results in under 300 milliseconds. Perplexity also released a Finance Search tool that returns structured data for public companies, covering quotes, financials, earnings reports, and analyst estimates, accessible through the same API interface as web search.
- When
- Published in the Perplexity API changelog, early October 2026.
- How it shifts discovery
- A 60 per cent price cut on the most heavily used agent tool in the API changes the cost ceiling for search-intensive agentic workflows. Any agent loop calling web_search multiple times per task now costs 60 per cent less at that step. Combined with the Fast Search default at sub-300-millisecond latency, both cost-per-search and speed-per-search improved in the same update cycle. Finance Search adds a different capability dimension: live structured financial data through the same API interface as general web search. For AI products targeting investment research, competitive market analysis, or any workflow requiring current public company data, this removes a separate tool integration and a distinct data vendor relationship. The direction is consistent: Perplexity is positioning its Agent API as the lower-cost, lower-latency retrieval layer for AI products that search at scale.
- Questions to ask
- The web_search price in Perplexity's Agent API dropped from $2.50 to $1.00 per 1,000 calls. For any AI product or agentic workflow currently calling web_search at scale, have we recalculated the revised monthly retrieval cost and confirmed whether the saving changes the cost comparison with alternative retrieval tools in the current stack?
- Finance Search returns structured public company data, financials, earnings, and analyst estimates through the same API interface as web search. For AI workflows that currently pull financial data through a separate tool or data vendor, does Finance Search cover the required data types at a price point that makes consolidation worthwhile?
- The Agent API fast preset now defaults to Fast Search at under 300 milliseconds. For AI products where retrieval latency constrains user experience or agentic task throughput, have we benchmarked Perplexity's fast preset against the current retrieval solution and documented the latency and cost tradeoff before the next development sprint?
- Sources
Key takeaways
What to walk away with this week
- 01
ChatGPT and Google AI Overviews can now both build functional tools inside their own interfaces: the two most-used AI discovery surfaces simultaneously completed the shift from answering questions to completing tasks, so audit every interactive conversion asset you own against what these platforms now generate for the same query this month.
- 02
Claude Haiku 5.5 and Perplexity's web_search price cut both arrived in the same week: the per-task cost of AI retrieval and structured reasoning is falling faster than most AI product budgets are set, so rebuild unit economics quarterly and do not plan 2027 AI spend against current prices.
- 03
Google's AI contribution pilot is the first formal mechanism to pay publishers for AI search content: the formula is opaque and payouts are wildly uneven, but this is the beginning of an AI search revenue model for content creators, so track which content categories attract the highest payouts and align your content strategy to how AI attribution appears to work in practice.
- 04
pplx-decider's Apache 2.0 open weights mean the decision layer above AI search results is now a commodity component: teams that build structured decision pipelines on open models this quarter gain an architecture advantage over those that remain dependent on closed-API general-purpose models for the same interpretation step.
- 05
The pattern across all five updates is the same: the AI layer is getting cheaper, faster, and more capable at resolving tasks inside its own interface, which makes brand presence in AI outputs a more urgent priority than ranking position in traditional results.
