AWS Adds GLM 5.3 to Bedrock at $1.68 Input and $5.28 Output

AWS launched GLM 5.3 on Bedrock on October 5

Amazon Web Services announced GLM 5.3 availability on Amazon Bedrock on October 5, 2026. This is a new managed route to Z.ai’s existing open-weight coding and agentic model, not a new GLM checkpoint. AWS says access is limited to eligible enterprise customers; teams without that entitlement must contact their AWS account team rather than assume the model will appear in every console.

The Bedrock model card marks the model active and documents text input and text output, a 1-million-token context window and up to 128,000 output tokens. The launch post positions the service for repository-scale coding, long-horizon agents and defensive security work. Those use cases are product positioning, not evidence that an unattended agent will remain correct or authorized over a million-token trajectory.

Requests must use a Global or US cross-Region profile

Applications must call global.zai.glm-5.3 or us.zai.glm-5.3. AWS lists zai.glm-5.3 as the base model ID, but explicitly says in-Region on-demand inference is unsupported. A request goes to a chosen source Region and Bedrock routes it for processing through the selected cross-Region inference profile.

That distinction belongs in the architecture record. Global CRIS gives AWS the broadest supported commercial-Region routing scope and carries the lower list price. The US profile keeps routing within its documented geography but is not the same as pinning processing to one Region. Verify the current destination set, account terms, logs and data requirements before sending regulated or location-restricted content.

Global Standard starts at $1.68 input and $5.28 output

The live Bedrock pricing table lists Global Standard rates per million tokens of $1.68 for input, $5.28 for output, $0.312 for cache reads and $2.10 for 30-minute cache writes. US CRIS costs $1.848 input, $5.808 output, $0.3432 cache read and $2.31 cache write—a 10% premium in each category.

AWS states that Priority pricing is 75% above Standard, while Flex and Batch are 50% below Standard. A tier multiplier is not a complete cost forecast: reasoning output, long agent trajectories, cache misses, repeated tool turns and retries can dominate a cheap input rate. Compare cost per accepted task and include service behavior, not only tokens from one successful call.

Four APIs do not mean four identical contracts

AWS supports the OpenAI-compatible Responses and Chat Completions APIs plus Bedrock InvokeModel and Converse. Its launch example uses the Responses API at the regional bedrock-runtime endpoint with global.zai.glm-5.3 and short-lived credentials generated from AWS CLI authentication. AWS recommends the OpenAI-compatible APIs for new applications because they expose the more complete feature set.

Treat each API surface as a separate integration until fixtures prove otherwise. Pin authentication, request fields, streaming event handling, structured output, function calls, reasoning controls, error mapping, retries and cancellation. A model-string replacement can still break parsers or authorization even when both providers describe an endpoint as OpenAI compatible.

Prompt caching has a price and an eligibility boundary

Implicit prompt caching is enabled by default. For explicit caching, AWS documents prompt_cache_options with cache breakpoints, a minimum of 1,024 tokens before each eligible checkpoint and retention of at least 30 minutes for this model. The pricing table separates cache reads from 30-minute cache writes, so both events must be visible in a cost model.

A breakpoint does not guarantee a hit. Stable prefixes, timing, routing and request construction still matter. During a trial, log input, cache-write, cache-read and output tokens separately; compare first-turn and repeated-turn latency; and test how prompt edits invalidate the prefix. Do not claim savings from the rate table without observed hit rates on the real conversation pattern.

The published model-size descriptions do not agree

AWS’s launch post calls GLM 5.3 a 753B-parameter mixture-of-experts model, matching the 753B label on Z.ai’s Hugging Face page. The live Bedrock model card instead says 744B total parameters and approximately 40B active per token. AccessAllGPT found no reconciliation on the reviewed pages, so procurement records should preserve both dated statements rather than silently choosing one.

The discrepancy does not change the Bedrock API price or profile ID, but it is a useful evidence-quality warning. Parameter count is not a performance result, and a hosted buyer does not operate those weights directly. Ask AWS or Z.ai to clarify the count if architecture scale is material to a disclosure, capacity comparison or technical assessment.

Coding and cyber scores remain vendor claims

Z.ai reports a 50% improvement over GLM 5.2 on its internal coding benchmark and publishes results including 28.3 on Terminal Bench 3.0, 66.9 on DeepSWE v1.1 and 84.5 on CyberGym. AWS repeats selected capability claims and demonstrates an authorized Strix security-testing workflow. AccessAllGPT did not reproduce any score or run that workflow; the numbers belong to the disclosed vendor harnesses and settings.

The model’s documented cyber capability raises the control requirement rather than granting authority. Any coding or security agent should run with scoped credentials, isolated execution, allowlisted targets, complete logs, step and spend ceilings, human approval before consequential action, and a kill switch. Only test systems you own or have explicit written permission to assess.

What builders and buyers should do next

Start by confirming account eligibility and selecting one cross-Region profile. Run a reversible shadow evaluation through the exact API and service tier intended for production. Freeze representative repository and tool-use tasks, set reasoning and output limits, and capture accepted outcomes, severe failures, latency, input, cache-write, cache-read and output tokens, tool calls, retries, refusals and cost.

Adopt only for task classes that clear predeclared quality, security, data-location and cost gates. Constrain the model to read-only analysis or patch proposals when execution authority is unnecessary. Wait when eligibility, routing or current pricing is not verified. Reject unattended consequential access without deterministic authorization, approval and rollback regardless of benchmark position.

Disclosures

AccessAllGPT received no AWS or Z.ai account, model access, credits, early information, briefing, demo, benchmark data, review or compensation for this article. Neither company sponsored, reviewed or endorsed it. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with Amazon Web Services, Z.ai, Hugging Face or Strix. Publication-wide relationships are listed on the disclosures page.