Strawberry agents now use 70% fewer credits. Update Strawberry and your existing plan should now last at least twice as long.
We just cut Strawberry’s credit use by 70%. Update Strawberry. Your current plan should now last at least twice as long! So, how did Arian and the agent team do it? Three main things: 1️⃣ We reduced edge-case cache misses Prompt caching lets devs reuse parts of a prompt instead of paying to process the same instructions and context again. A complex agent harness creates a large surface area for cache misses. Adding/removing tools, dynamically changes to prompts, memory updates, etc. can all invalidate cached prompts. These edge cases are very hard to discover in in-house evals, so we improved our anonymous telemetry around real-world cache behaviour. We found many common miss patterns, and patched them. With a nice Dash0 dashboard, it is now easy to see which situations cause cache misses. If you are building agents and not measuring cache hit rates, you have to look into this. 2️⃣ We switched our base model to GPT-5.6 Terra In our evals, Terra gives us Opus 4.8-level performance at roughly half the model cost. Its tokenizer is also much more efficient. For the average Strawberry prompt, Terra uses roughly half as many input tokens as Anthropic’s models. Your Strawberry agents now use cheaper input tokens, and fewer of them. 3️⃣ We made our agent harness more context-efficient Agent harnesses can accumulate absurd amounts of input context. Reading a 1,000-row sheet in full can push more than 100,000 input tokens into a single prompt. Our new harness limits large tool results, preserves the useful parts, and lets the agent surgically fetch more context when needed. We improved memory search using the same principle. Strawberry's memory system can contain thousands of memory files covering candidates, meetings, customer interviews, and much more. We now give the agent a better overview of the file tree, helping it search effectively and open only the relevant files. Most agent cost problems are context-engineering problems. Happy Monday :)