This week delivered one direct model-pricing change, two new ways to buy agentic AI, a cloud procurement route for controlled cyber models, and an open-model announcement that is not yet priceable. The live table, chart, and calculator above read current API rates from our maintained dataset; they do not treat subscription bundles or unreleased weights as token-priced models.
| Story | What changed | Buyer impact |
|---|---|---|
| Grok 4.6 | xAI launched a frontier coding and agent model with a new cache rate | Benchmark repeated long prompts before migrating from Grok 4.5 |
| Daybreak on AWS | Approved teams can buy Blue and Red through Amazon Bedrock | AWS governance may justify the provider uplift, but it is not the cheaper route |
| Claude Code Auto mode | Anthropic removed classifier overhead for most paid users | Permission checks use less allowance, while longer runs still need budgets |
| Grok Bot | xAI bundled persistent cloud-computer agents into eligible plans | Subscription access is only the floor when usage becomes metered |
| Genesis-Science-1 | DOE opened development of a scientific workflow model | Wait for weights, terms, benchmarks, and infrastructure data before budgeting |
Grok 4.6 changed cache economics
xAI released Grok 4.6 for coding, agentic tasks, and knowledge work with text-and-image input and a 500,000-token context window. Its standard and long-context input and output rates match Grok 4.5, but cached input is more expensive. Prompts that cross the long-context threshold also move the whole request onto a higher rate tier.
That makes the launch attractive to cache-light frontier workloads and less automatic for repository agents that repeatedly send large system prompts, tool schemas, maps, or conversation histories. Canary the pinned model, measure real cache hits, and compare cost per accepted task at each reasoning level. See the full Grok 4.6 pricing analysis and live xAI API pricing.
Daybreak reached Amazon Bedrock
OpenAI made Daybreak Blue and Daybreak Red available to approved security teams through Amazon Bedrock. Blue provides GPT-5.6 Sol with safeguards adapted for authorized defensive work; Red provides the purpose-trained GPT-5.6 Cyber model for tightly governed vulnerability and exploit validation.
AWS’s listed token rates carry an uplift over direct OpenAI short-context rates. The Bedrock route can still win for organizations that value centralized identity, audit, cloud commitments, and procurement, but teams should compare cost per validated and fixed vulnerability—not cost per prompt. Read the Daybreak on AWS pricing breakdown and current OpenAI pricing.
Coding agents shifted from prompts to autonomy
Anthropic stopped charging Pro, Max, and Team users for Claude Code Auto mode’s classifier-token overhead and will make Auto mode the default for those plans on August 14 unless another permission mode is pinned. Base subscription and Claude API rates did not change. Fewer approval interruptions may improve throughput, but they can also let a weak loop consume more tokens before a human notices.
xAI separately launched Grok Bot: persistent named agents that share one cloud computer per user and can operate browsers, terminals, files, connectors, and scheduled routines. There is no standalone Bot rate; access is bundled into eligible xAI or Cursor subscriptions, with metered usage available after included allowances. Compare the Claude Code Auto mode analysis with the Grok Bot pricing guide.
For a coding-only alternative with a separate quota, test the Z.ai coding plan on the same repository task and compare accepted changes, retries, and review time.
Affiliate disclosure: we may earn a commission from the sponsored link above. It does not affect our analysis.
Genesis-Science-1 is not priceable yet
The U.S. Department of Energy named Genesis-Science-1 as the first planned model in its Genesis Open Models Initiative. The goal is an open-weight model that works through scientific code, data, tools, simulations, failures, and recovery—not merely a research chatbot.
DOE opened contribution rounds but has not released weights, a license, architecture, hosted endpoint, or API price. Research teams can prepare realistic workflow evaluations now, but should keep current systems in production until a pinned release has measured accelerator-hours, task success, retries, and expert-review cost. Our Genesis-Science-1 analysis separates open access from deployment economics.
What buyers should do now
Treat model access, agent autonomy, and deployment route as separate cost decisions. Replay cache-heavy workloads before moving to Grok 4.6, price Daybreak with governance and analyst labor included, cap every autonomous coding run, and refuse to budget unreleased weights from headline claims.
Use the AI token calculator for published model rates, then add tools, retries, review, infrastructure, and subscription allowances. This week’s common lesson is simple: the cheapest rate card does not identify the lowest cost per accepted result.
Sources: xAI’s Grok 4.6 release notes and Grok Bot announcement, OpenAI’s Daybreak on AWS announcement, Anthropic’s Auto mode announcement, and DOE’s Genesis Open Models Initiative.