gpt-6.1 terra on Bedrock: Speed, Price and the Fine Print

gpt-6.1 terra became generally available on Amazon Bedrock on October 26, 2026, and it is the kind of release that changes procurement math more than headlines. AWS positions it as the GPT-6 family option balanced for capability, speed and cost on everyday workloads: visual reasoning, computer use, and improved factuality and natural writing over GPT-5.6 Terra, with token prices at least 50 percent lower than GPT-6.1 Sol and GPT-5.6 Terra. For teams already building agents on Bedrock, the practical questions are the ones this post answers from the primary documentation: what the benchmark claims actually say, what it costs at short and long context, which APIs and features work, and where the limits are. Source

What GPT-6.1 Terra is

Per the AWS announcement, GPT-6.1 Terra is the newest model in OpenAI's Terra line, aimed at complex workflows that need detailed visual reasoning and accurate computer use. It accepts text and image inputs and produces text output, with a context window of 1,050,000 tokens and a maximum output of 128,000 tokens according to the Bedrock model card. AWS says you can call it directly through the Bedrock APIs or configure ChatGPT Work and Codex to use the model on Bedrock, and it pitches the model at work like drafting proposals, interpreting technical drawings, investigating codebases and building agents that operate across business systems. The lifecycle commitments are unusually specific for a frontier model: end of life no sooner than October 26, 2027, with a legacy period of at least six months, which matters for teams that have been burned by fast model deprecations. Source

The benchmark claims, with the right framing

The headline number comes from OpenAI's own evaluation called Agents' Last Exam: GPT-6.1 Terra reportedly scores 15 percentage points higher than GPT-5.6 Terra, performs on par with GPT-6 Astra, and completes tasks more than three times faster than both GPT-6 Astra and GPT-6.1 Sol. Frame these correctly. This is a vendor-run evaluation on a vendor-named exam, repeated in AWS's marketing, and we did not find independent verification of the scores or the speed claims. The claims are also comparisons within OpenAI's own lineup, not against Claude, Gemini or the open models, so they say nothing about whether Terra is the right model, only where OpenAI positions it. If the par-with-Astra and 3x-faster claims hold in your own testing, the cost implication is significant, which is exactly why you should test on your tasks before moving traffic. Source

The ChatGPT Work and Codex angle

One line in the announcement deserves more attention than it got: you can configure ChatGPT Work and Codex to use GPT-6.1 Terra on Bedrock. For organizations that prefer to keep model invocation traffic inside AWS for governance, this addresses a common objection. On the documented configuration, the model invocations made through Bedrock are auditable through the same AWS controls you already use for access and logging, and AWS pitches its established controls for securing workloads, governing access and auditing invocations; the announcement does not describe what logging applies to the rest of a ChatGPT Work or Codex session. For AgentOps teams, an auditable invocation path can remove one common blocker in a security review, though whether it satisfies a given review depends on that organization's requirements, and it follows the same governance-first pattern we recommend when teams evaluate agent platforms. Source

Pricing: the two context tiers

Bedrock's pricing for GPT-6.1 Terra is Standard tier only; Priority and Flex tiers are not supported, and neither is Reserved throughput. For short context, defined as 272K input tokens or fewer, In-Region and Geo cross-Region inference cost $1.10 per million input tokens and $5.50 per million output tokens, while Global cross-Region inference is $1.00 and $5.00. The 30-minute cache write costs $1.375 per million tokens and cache reads drop to $0.055, which makes prompt caching the obvious lever for repeated system prompts. Above 272K input tokens, prices step up sharply: $2.20 input and $8.25 output In-Region, or $2.00 and $7.50 on Global CRIS. That long-context surcharge, roughly double on input, is the number to model before you design an agent that stuffs whole codebases into every call; retrieval discipline still pays. Commercial In-Region prices already include the 10 percent fee over OpenAI's rates. Source

APIs, endpoints and feature support

The model card draws a clear split between endpoints. On bedrock-runtime, the model works with the Responses, Chat Completions, Converse and Invoke APIs; on the bedrock-mantle endpoint, only Responses and Chat Completions, and both live under an /openai/v1 base path, with AWS recommending bedrock-runtime for new applications. Feature support on bedrock-runtime includes response streaming, structured outputs with JSON Schema, reasoning summaries, implicit prompt caching across APIs and explicit caching on the Responses API, and Guardrails through the Converse API only. Just as important is what is missing: no server-side tool use, no intelligent prompt routing, no token counting endpoint and no watermarking. No server-side tool use means your agent loop runs client-side as it would with a self-managed stack, which we see as the right default anyway for teams that want portable tool permission logic, the subject of our AI agent tool permissions guide. Source

Regions and access patterns

Regional availability is worth reading carefully because it differs by endpoint and inference type. The bedrock-mantle endpoint supports the model in us-east-1 only. On bedrock-runtime, In-Region inference is not available in the listed US regions, but Geo and Global cross-Region inference profiles work from us-east-1, us-east-2, us-west-1, us-west-2 and ca-central-1, with programmatic requests naming the cross-Region inference ID us.openai.gpt-6.1-terra as the model. GovCloud regions are not supported. For data-residency-sensitive workloads, the Geo versus Global CRIS choice is a real decision: Global is cheaper but routes across AWS's global infrastructure, so check your compliance requirements before defaulting to the cheaper tier. Source

Where it fits, in our assessment

This section is our read, separate from the sources. GPT-6.1 Terra lands one day after Meituan's longcat-2.5 technical post, and the two releases frame the same market from opposite ends: OpenAI pushing a cheaper, faster workhorse through AWS's enterprise channel, and Chinese labs publishing agentic-training research to close the capability gap. For founders and platform teams, our advice is mechanical. First, if you run agents on Bedrock today, put GPT-6.1 Terra into your eval harness against your current model on your own tasks; the vendor claims point at real value but only your workload can confirm them, and our AI MVP tech stack post covers how to structure that comparison. Second, design around the 272K pricing cliff rather than discovering it on an invoice. Third, treat the one-year EOL commitment as planning cover, not a reason to skip abstraction: keep your model calls behind an interface so the next Terra costs a config change, not a rewrite. Teams that skip that abstraction tend to discover its cost at the worst moment, mid-incident or mid-migration, and the pattern of every model cycle so far suggests another cheaper workhorse is always one announcement away. Source