Claude 4 API: Tool Use Streaming & Prompt Caching vs Together AI DeepSeek R2 Distilled Serverless Inference

“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: OpenAI-compatible serverless inference endpoint for distilled reasoning models, no infra to manage. The DX bet Together AI made is correct — zero-config model access with standard chat completions API means you swap one base URL and one model string and you're calling DeepSeek R2 distilled from existing code. The 32B at $0.80/M tokens is the real story: that's sub-dollar-per-million for a model that punches well above its weight class on reasoning benchmarks. The weekend alternative is self-hosting on RunPod or Modal, which works but adds cold-start latency, VRAM management headaches, and ops overhead that Together simply removes. Ship this if you're building anything that needs cheap chain-of-thought reasoning without the frontier model bill.”

“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 Fireworks AI, Groq, and Replicate running the same or similar distilled checkpoints — so Together is not selling exclusivity, they're selling reliability and price. The scenario where this breaks is high-concurrency production workloads where serverless cold-start variance becomes a latency SLA problem; Together's serverless tier has no guaranteed throughput contracts in the base offering. What kills this in 12 months is not a competitor but the underlying model provider: if DeepSeek ships R3 distills that are 2x better at the same cost, this specific offering goes stale and Together has to scramble to re-serve. That said, Together's track record of being early on new model availability is the actual moat here — they've consistently been first or second to serve hot open-weight checkpoints, and that speed-to-availability is worth paying for if you're iterating fast.”

“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 Together AI is betting on: by 2027, the majority of production LLM inference will run on open-weight distilled models, not frontier APIs, because the quality gap closes faster than the price gap opens. That's a falsifiable and plausible claim — the DeepSeek R1 distillation story already validated it at the 7B-32B range. The dependency that has to hold is that distillation techniques keep pace with frontier capability jumps, which is not guaranteed if frontier labs accelerate architectural innovation faster than distillation pipelines can follow. The second-order effect that's underappreciated: cheap reasoning inference at this scale shifts power from model labs to inference infrastructure providers — Together, Fireworks, Groq become the AWS to the model labs' hardware vendors. Together is on-time to this trend, not early, but their execution on catalog breadth means they're well-positioned if the trend accelerates.”

“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 any developer or startup running LLM inference who currently pays OpenAI or Anthropic rates for reasoning tasks that don't require frontier-model quality — that's a real and large budget line item. The pricing architecture is usage-based and scales directly with value delivered, which is the right structure for inference. The moat question is harder: Together's defensibility is not the models (open weights, anyone can serve them) but latency, reliability, and the breadth of the model catalog creating switching friction once you've standardized your inference client on their SDK. The existential risk is that this is fundamentally a margin business on commodity compute, and Cloudflare Workers AI, AWS Bedrock, and Google Vertex are all moving to serve the same checkpoints at infrastructure-subsidized prices. Together needs to win on speed-to-new-models and developer experience before the hyperscalers catch up on catalog breadth, and so far they're doing it.”