AWS in October: Haiku 5.5 75% Cheaper, Model Fallback and a Sandbox for Agents

Running a single agent in production is easy. It gets harder when there are dozens: models hit rate limits, bills grow, and an agent suddenly wanders somewhere it wasn't invited. AWS's October digest is almost entirely about these problems. Below is what a practicing developer or product manager should take note of. The figures are based on statements from AWS and vendors; the material includes no independent measurements.

Models for the Task and the Budget

The main idea of the month: don't route everything through a single flagship model. Claude Haiku 5.5 is built for subagents and high-volume tasks. According to Anthropic, it is roughly 75% cheaper than Haiku 4.5 for most tasks. It is the first Haiku with an effort setting: you can choose how much to pay for reasoning depth in a given request.

The typical setup looks like this: Claude Opus 5.5 plans the work, while Haiku executes well-defined chunks: code, tool calls, browser automation. Running many agents in parallel stays inexpensive.

GPT-6.1 Terra is now generally available on Bedrock. On OpenAI's internal Agents' Last Exam benchmark, it is comparable to GPT-6 Astra but solves tasks more than three times faster. Input and output token prices are at least 50% lower than for GPT-6.1 Sol and GPT-5.6 Terra. It offers a context window of up to 1M tokens, with text and image support.

For voice, there is Amazon Nova 2.5 Sonic: the model reasons better and calls tools more accurately, at the same price. A support agent can find an order, check refund eligibility, and process the refund in a single conversation.

Reliability: Fallback, Workflows and JSON

The most useful addition for production turned out to be AgentCore Gateway. It now accepts Bedrock InvokeModel calls and supports model fallback. If the primary model is rate-limited or unavailable, the request goes to the next model or provider in the chain you define. A circuit breaker skips endpoints that are down. Existing code built on the Bedrock SDK just needs to be pointed at the gateway; there's no need to write retries in every client.

The second important item: Step Functions as agent tools. A business process, such as a refund, stays in a workflow you already operate. The agent decides when to trigger it, while the refund eligibility check, approval, and payout run strictly through your steps rather than however the model interpreted the prompt.

Structured outputs are now available for all Claude models on Bedrock: JSON that follows a schema, with no hand-written parsing or retries. Web search gained domain filters (up to 100 allowed and up to 100 blocked) and region binding. Guardrails checks that the response doesn't diverge from the sources. The new Standard tier of checks costs the same as Classic, but covers conversational and reasoning responses and more languages.

Money and Strands

AgentCore Memory is now billed by data volume rather than a fixed price per event. Scenarios with small payloads benefit the most. A TTL can be set for long-term memory. In Cost Explorer and Budgets, Bedrock spend is broken down by model, provider, inference type, and feature, at no charge. A budget alert can be set on a single model.

In Strands, the open-source agent toolkit, the headline is Strands Box in developer preview. It is a local sandbox with OS-level isolation and per-tool policies. They apply to direct calls, to calls through MCP, and to code generation. Strands Evals gained red teaming: prompt injections and jailbreaks are run against your agent, and you can see at which message it broke. Strands Decider, a small model based on Gemma 4, chooses among given options instead of generating text. In one internal Amazon ticketing system, the estimated cost per ticket dropped by about 98% and response time by 78%.

What to Do With This

This is a vendor digest, so the numbers should be verified on your own workloads. Most of the features are tied to the AWS ecosystem. But the set of approaches is universal. Split planning and execution between an expensive and a cheap model. Move critical processes out of prompts and into deterministic workflows. Set up fallback before the first incident. Turn on the per-model cost breakdown before the bill becomes a surprise.