Claude 4 API: Tool Use Streaming & Prompt Caching vs Perplexity Deep Research API

“The primitive here is clean: incremental tool-call deltas over SSE, and a cache-control header you attach to prompt segments to pin them server-side. The DX bet is that complexity lives in the HTTP layer, not in a new SDK abstraction — you opt in per-request, no new mental model required. The moment of truth is calling `stream=true` on a tool-use request and watching partial JSON arguments arrive before the model finishes thinking, which actually matters for agent loops where you want to dispatch work early. This is not a weekend-script replacement — implementing correct incremental JSON parsing for partial tool arguments plus a reliable distributed cache with 2M token capacity is a real engineering problem Anthropic has solved for you. The specific decision that earns the ship: cache invalidation is explicit and cache hits are reflected in the usage object, so you can actually measure what you're saving instead of guessing.”

“The primitive here is clean: one API call returns a fully cited, multi-step research synthesis instead of raw search results you have to reassemble yourself. The DX bet is that developers would rather pay per-request than build query decomposition, iterative retrieval, and deduplication logic on top of a search API — and that's actually a reasonable bet for most product teams. The 10-minute moment of truth is solid: get an API key, POST a query, get back structured citations and a synthesized answer. The weekend alternative would be stitching together a search API, chunking strategy, and an LLM into a loop — achievable but genuinely annoying, especially for fresh web content. What earns the ship is that this isn't a wrapper around a single endpoint — it's exposing a multi-hop retrieval pipeline that would take real engineering hours to replicate at comparable quality.”

“Direct competitors are OpenAI's cached completions and Google's context caching in Gemini 1.5 — both shipping for months — so Anthropic is catching up, not leading. The specific scenario where this breaks: cache hit rates depend entirely on prompt structure, and developers who dynamically compose system prompts (inserting user-specific context at the top) will see near-zero cache utilization and pay full price while assuming they're saving money. The prediction: this feature doesn't get killed — it becomes table stakes infrastructure and Anthropic wins by having the largest cache window (2M vs. competitors' current limits). What would have to be true for me to be wrong: OpenAI ships a 10M token cache window before Anthropic's ecosystem matures, commoditizing the advantage. Still a ship because the streaming tool-use delta is genuinely differentiated — no competitor has clean partial-argument streaming for tool calls yet, and that changes agent loop architecture in ways that matter.”

“Direct competitors are OpenAI's own web search tool in the Responses API, Exa's research endpoints, and anyone building on top of Tavily or Brave Search with an LLM loop — so the market is genuinely crowded. Where Perplexity has a real edge is that Deep Research is not one LLM call plus search; it's iterative, it self-directs, and the citation quality is demonstrably better than naive RAG. It breaks at scale: high-frequency, time-sensitive queries will get rate-limited and the per-request cost will hurt anyone building a high-volume product without careful caching. What kills this in 12 months is that OpenAI ships a comparable multi-step research endpoint natively in the Responses API and undercuts on price — that's the most plausible outcome. What earns the ship anyway is that Perplexity is genuinely ahead on research quality today, and shipping into that window while it exists is a legitimate product strategy.”

“The thesis this bets on: by 2027, the dominant AI application architecture is a persistent agent with a large, stable context (tools, memory, instructions) that gets reused across thousands of user interactions — making context I/O cost the primary unit economics lever, not generation cost. The dependency that has to hold: agents don't collapse back to stateless chatbots, and context windows keep growing faster than per-token prices fall. The second-order effect nobody's talking about: prompt caching at 2M tokens makes it economically viable to give every enterprise user a fully-loaded, role-specific agent context at request time — which shifts competitive differentiation from 'who has the best model' to 'who has the best cached context corpus,' effectively making knowledge curation the new moat. This tool is riding the trend of context-window expansion-as-infrastructure, and it's on-time, not early — but the streaming tool-use primitive is ahead of the curve on agent loop efficiency. The future state where this is infrastructure: every production agentic system has a cache manifest the same way it has a CDN config.”

“The thesis this API bets on: in 2-3 years, most knowledge-work applications will need live web synthesis as a primitive, not a feature they build themselves — the same way they stopped building their own payment infrastructure. That's falsifiable: it fails if model providers commoditize retrieval-augmented generation to the point where there's no differentiated value in a managed research pipeline. The second-order effect that matters here isn't the direct API revenue — it's that Perplexity gets embedded in the output layer of dozens of third-party products, which compounds their training signal and usage data. The specific trend line is the shift from search-as-lookup to search-as-synthesis, and Perplexity is genuinely on-time here while most competitors are still early. The future state where this is infrastructure is every B2B SaaS product embedding a research tab — not because they want to, but because not having one becomes a competitive disadvantage.”

“The buyer is the engineering team at any company running Claude in production with long system prompts or multi-step agents — this comes out of the AI infrastructure budget, not a new budget line, which means no procurement friction. The pricing architecture is sound: cache reads at ~90% discount means the savings are real and measurable in the first billing cycle, which creates immediate retention — developers who restructure prompts to maximize cache hits are now architecturally coupled to Anthropic's caching implementation. The moat question is the honest one: this is infrastructure that OpenAI and Google will match, so the defensible position isn't the feature itself but the ecosystem of developers who've restructured their codebases around it. What survives a 10x model price drop: the streaming tool-use architecture, because that's about latency, not cost. The specific business decision that makes this viable is pricing cache reads as a separate SKU — it lets Anthropic capture value from high-volume production workloads without losing price-sensitive experimenters.”

“The buyer is a product team at a B2B SaaS or research tool company that has a line item for API infrastructure — this comes from engineering or product budget, not a standalone tool budget. Pricing at pay-per-use aligns with value but creates a land-mine for consumer-facing apps where one viral feature can spike costs by an order of magnitude; any serious team will need rate-limiting and cost caps before shipping to end users. The moat is real but narrow: Perplexity's citation quality and iterative research pipeline are ahead of commodity alternatives today, but this is a capability moat, not a data or distribution moat, which means it erodes as frontier model providers close the gap. The business survives if Perplexity becomes the default research infrastructure layer for the developer ecosystem before OpenAI or Anthropic ship a comparable managed endpoint — that's a plausible 18-month window and they're moving into it. Ships because the unit economics work for mid-volume use cases and the wedge into developer workflows is real.”