Claude 4 API: Tool Use Streaming & Prompt Caching vs FlashInfer 2.0

“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 a CUDA kernel library for attention computation and KV-cache management — not a platform, not a wrapper, an actual low-level building block you can drop into vLLM or SGLang. The DX bet is correctness and composability over abstraction: they expose the knobs (speculative decoding thresholds, LoRA batching configs) without hiding them behind a config YAML that pretends the complexity doesn't exist. The moment of truth is swapping in the FlashInfer attention backend in an existing serving stack, and from what the repo shows, that's genuinely a few lines. The 40% latency claim needs a methodology cite — they show specific token generation benchmarks on H100s with prefill/decode separation, which is at least a real number attached to a real setup, not a vibe. This is infrastructure that a competent team could not replicate in a weekend; the CUDA work is deep and the speculative decoding integration is non-trivial. Ships because the craft is demonstrably in the kernels, not the landing page.”

“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.”

“Category is LLM inference optimization, direct competitors are FlashAttention-3, vLLM's built-in attention kernels, and NVIDIA's TensorRT-LLM — none of which are sleeping. The 40% latency claim is real in a narrow regime: it applies to specific decode-heavy workloads on Hopper-generation GPUs with prefill-decode disaggregation; swap in an A100 cluster doing long-context prefill and the number shrinks. What kills this in 12 months is not a competitor — it's NVIDIA shipping optimized kernels directly into cuDNN or the next-generation attention primitives landing in TensorRT-LLM, at which point the delta collapses. What earns the ship anyway: multi-LoRA batching at scale is a genuinely underserved problem that the big players haven't prioritized, and Together AI has production traffic to validate these claims against real workloads, not synthetic benchmarks. The open-source release is credible signal that they're playing for ecosystem, not just headlines.”

“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 here is specific and falsifiable: inference compute will remain the dominant cost in LLM deployment for at least the next three years, and kernel-level optimization will continue to yield meaningful gains even as hardware scales. What has to go right is that the prefill-decode disaggregation architecture becomes the dominant serving pattern — if monolithic batching stays standard, FlashInfer's architectural assumptions become a liability rather than an asset. The second-order effect that matters most isn't latency reduction for Together AI's own platform — it's that cheap, reliable multi-LoRA serving changes the economics of fine-tuning. If you can serve 50 LoRA adapters off one base model at acceptable latency, the cost of domain-specific fine-tuning drops by an order of magnitude, which shifts power toward the fine-tuning layer and away from base model providers. FlashInfer is riding the prefill-decode disaggregation trend, and it's on-time rather than early — vLLM and SGLang have already moved this direction, which means the ecosystem is ready to absorb this rather than resist it.”

“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 here is infrastructure engineers at companies running self-hosted LLM inference at scale — a real buyer with a real budget (GPU compute costs), not a vague enterprise persona. The open-source release is a distribution play, not a charity: Together AI captures value through their managed inference platform, where FlashInfer improvements directly reduce their per-token compute cost and become a credible differentiator in a market where Fireworks, Groq, and Anyscale compete on latency benchmarks. The moat question is the hard one — open-sourcing the kernel library means competitors can adopt it too, so the defensibility is execution velocity and production integration depth, not the code itself. What happens when NVIDIA ships this natively is the real stress test, and the honest answer is that Together AI's moat shifts entirely to their managed platform and the workflow integrations built on top of it. Still a ship because the business logic is coherent: they're using open source to build pipeline credibility while monetizing on the managed layer, which is a proven playbook.”