Claude 4 API: Tool Use Streaming & Prompt Caching vs Together AI Llama 3.3 Fine-Tuning 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: submit a dataset, get back a LoRA adapter, deploy it — no CUDA drivers, no FSDP config, no sacred Hugging Face trainer incantations. The DX bet is to hide all the distributed training complexity behind a single API call, which is the right call for 80% of fine-tuning use cases. The auto-eval runs are a genuinely useful addition — getting a held-out eval without writing your own harness is the kind of thing that saves a Tuesday afternoon. My one gripe: the 'one-click deployment' language is landing-page speak until I see the actual API surface for versioning and rollback. If that's solid, this is a legitimate skip-the-weekend-script win; if it's a button in a dashboard with no programmatic control, it's half a tool.”

“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.”

“The direct competitor is Modal plus Axolotl, or just calling the OpenAI fine-tuning API — and that comparison is where Together has to win. They do have a credible answer: Llama 3.3 is open-weight and OpenAI won't fine-tune it for you, so if you want this specific model, Together is a real option rather than a convenience wrapper. The scenario where this breaks is at scale: teams with large proprietary datasets and strict data residency requirements will hit contractual blockers before they hit a technical one. The 12-month kill scenario is that Meta ships a hosted fine-tuning offering tied to its own inference cloud, or Groq and Fireworks match this and compete on price, squeezing Together's margin to zero on a commodity service. What would have to be true for me to be wrong: Together builds enough workflow lock-in through evals, versioning, and deployment that switching cost exceeds the price delta.”

“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: within 2-3 years, fine-tuning open-weight models becomes as routine as calling a hosted API today — the infrastructure friction is the only thing stopping most teams from doing it. That's a falsifiable and plausible bet; the trend line is the declining cost of LoRA training on commodity hardware, and Together is early-to-on-time, not late. The second-order effect that matters isn't that teams customize Llama — it's that model customization stops being a specialized MLOps discipline and becomes a product feature anyone can ship, which shifts power away from model providers with closed APIs toward whoever controls the fine-tuning workflow layer. The dependency that has to hold: open-weight models must remain competitive with closed frontier models for the tasks where fine-tuning provides the edge. If GPT-5 or Gemini 2.x make fine-tuning irrelevant by being few-shot-capable enough for every use case, the whole thesis collapses.”

“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 an ML engineer at a mid-size tech company whose team doesn't want to manage GPU clusters — that's a real person with a real budget line. But the moat here is essentially zero: this is compute arbitrage plus a thin API wrapper, and every inference provider with spare H100s can ship the same thing in a quarter. The pricing scales with training compute, which means Together's margin collapses exactly when the customer is getting the most value — high-volume fine-tuning jobs. What would need to change: Together would need to build proprietary eval infrastructure, dataset tooling, or model versioning deep enough that the workflow lock-in survives a 40% price cut from a competitor. Right now it's a good product that isn't a good business.”