Four frontier models shipped in one week. The practical split is access, cost, and control. (2026)

Between September 1 and September 3, Anthropic, Google, Meta, and OpenAI each put a new frontier model or model family in front of developers. The releases overlap on coding and long-running agent work, but they make different trade-offs: GPT-6 Astra buys higher capability with a much higher API price, Claude Fable 5.1 makes agentic work cheaper through cache pricing and effort controls, Gemini 3.8 Flash keeps a 1M-token input limit at Flash prices, and Muse Spark 1.3 focuses on agents that plan, recover, and ask for help.

The week matters because a first evaluation can now answer a more specific question than "which model is smartest?" A team can test whether it needs Astra's token efficiency, Fable's coding quality, Gemini's broad tool surface, or Muse's long-horizon behavior before it changes a production route.

Model	Builder and ship date	Access and scale	Strongest first test	Action window	Main limit
GPT-6 Astra	OpenAI, Sep. 3, 2026 1	API, ChatGPT, Azure, and Bedrock; $10 input / $50 output per million tokens 2	Long coding and computer-use tasks where fewer tokens can offset a higher rate	Start now if a workflow is limited by long runs or output volume	Production safeguards can pause legitimate work, and cyber claims differ between research and deployed settings 2
Claude Fable 5.1 / Mythos 5.1	Anthropic, Sep. 1, 2026 3	Fable is broadly available; Mythos is restricted to trusted cyber and life-science access programs; Fable costs $10 / $50 per million tokens 3	Long coding, research, and document workflows with repeated context	Start now for Fable; apply only if Mythos's restricted capabilities fit the work	Safeguards can redirect penetration testing, exploit generation, and binary scanning to other models 3
Gemini 3.8 Flash / Flash Cyber	Google, Sep. 2, 2026 4	API and Google developer products; 1,048,576-token input limit; $0.75 / $3.75 per million tokens through Dec. 31, 2026 4 5	Multimodal, tool-using agents over large document sets	Start now for high-volume multimodal or long-context work	`minimal` thinking is unsupported; Flash Cyber is restricted to trusted defenders 4 5
Muse Spark 1.3	Meta, Sep. 2, 2026 6	Muse Code and Meta Model API; price and context limit are not stated in the launch post	A single-thread agent that must plan, use tools, recover, and ask clarifying questions	Start now if agent reliability matters more than a published leaderboard score	Meta gives qualitative gains and tool-use reductions, without a numeric scorecard in the post 6

GPT-6 Astra: fewer tokens, higher bill

OpenAI released GPT-6 Astra on September 3. The model is available through ChatGPT, the API, Microsoft Azure, and AWS Bedrock. OpenAI lists $10 per million input tokens and $50 per million output tokens, compared with $4 and $20 for GPT-5.6 Sol. The model has a 1,050,000-token context window and a 128,000-token maximum output according to the launch material. 1 2

Astra's intended job is broad computer work: browser tasks, software engineering, scientific analysis, cybersecurity, and professional documents. OpenAI reports 72.6% on OSWorld 2.0 at about 40 minutes per task, compared with 65.7% for GPT-5.6 Sol at about 75 minutes. OpenAI also reports 57.9% on Terminal-Bench 4.0 and 88.0% on SRE-Bench in a single attempt. These are OpenAI's evaluations, and the launch page says some results use research or API configurations rather than the exact production ChatGPT or Codex setup. 2

Independent testing gives the price trade-off a sharper shape. Artificial Analysis scores Astra at 61 on its Intelligence Index, equal to GPT-5.6 Sol and five points behind Claude Fable 5.1 at maximum effort. Astra scores 67 on the Coding Agent Index, roughly matching Fable 5 and Opus 5 in their coding environments, while Fable 5.1 in Claude Code scores 70. Artificial Analysis also finds that Astra uses about one-third as many tokens as GPT-5.6 Sol in its coding harness, but the 2.5-times price increase leaves Astra about 75% more expensive per task on its Intelligence Index. 7

Astra belongs in a test when a workflow spends heavily on long answers, repeated planning, or computer actions. Run the same tasks against GPT-5.6 Sol and one close competitor. Track success, wall-clock time, output tokens, retries, and the number of human approvals. The model's value depends on whether its lower token use outweighs its higher token price on the reader's workload.

Cybersecurity needs a separate test. OpenAI says Astra crossed the Critical cybersecurity capability threshold in its Preparedness Framework. The launch describes strong research results without production safeguards, while the deployed model refuses more advanced requests such as proof-of-concept exploit creation. OpenAI also says monitoring can pause or stop work. A security team should therefore test defensive review, patch generation, and false refusals in a controlled environment rather than infer production behavior from the headline benchmark. 2

Claude Fable 5.1 and Mythos 5.1: one model, two access regimes

Anthropic released Claude Fable 5.1 and Claude Mythos 5.1 as the same underlying model with different safeguards. Fable 5.1 is generally available in Claude, Claude Code, Claude Cowork, the Claude API, Amazon Web Services, Google Cloud, and Microsoft Azure. Mythos 5.1 is limited to trusted access programs for cybersecurity and life-science organizations. 3

Fable 5.1 keeps the $10 input and $50 output price of the current premium tier, while cache reads fall to $0.25 per million tokens. Anthropic estimates a 25% lower cost for typical token-billed work and up to 45% lower cost for highly agentic, context-heavy work. Effort settings also change the bill: Fable 5.1 defaults to High in Claude Code and Medium in Claude Cowork and Claude.ai. 3

Anthropic reports 55.8% on Terminal-Bench 4.0 for Fable 5.1, 73.4% on CursorBench 3.2.0, and 1853 on GDPval-AA v2. The same table reports 37.3% for GPT-5.6 Sol on Terminal-Bench 4.0 and 67.2% on CursorBench. The benchmark notes matter: Fable 5.1 ran with production safeguards, and safeguard interventions produced zero scores on some OSWorld tasks. Terminal-Bench-Science also carries a standard error of roughly 3.5 to 4.5 points, so small gaps deserve little weight. 3

Mythos 5.1 is a different access decision from Fable 5.1. Anthropic reports a 60.9% Terminal-Bench 4.0 score for Mythos 5.1, but the model is available only through trusted programs. The restricted route matters to organizations that need broader cyber or biology capability under a monitored access arrangement. Most teams should evaluate Fable 5.1 first, especially for code review, research, and unattended document work.

Gemini 3.8 Flash: long context at a Flash price

Google released Gemini 3.8 Flash and Gemini 3.8 Flash Cyber on September 2. Gemini 3.8 Flash accepts text, images, video, audio, and PDFs, and supports caching, code execution, computer use in preview, file search, function calling, search grounding, structured outputs, and three thinking levels: Low, Medium, and High. Google's model documentation lists a 1,048,576-token input limit and a 65,536-token output limit. 4 5

The introductory API price is $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026. Google says the price will become $1.50 and $7.50 from January 1, 2027. Google reports a 54.9% score on HLE-Verified and describes gains on software engineering, agentic tasks, and long-horizon workflows. 4

Gemini 3.8 Flash is the low-friction candidate for teams that need one API to read mixed media, call tools, and search across very large inputs. The first evaluation should use a fixed set of long documents, screenshots, and tool calls. Measure retrieval accuracy at 128K and 512K tokens before assuming that the advertised ceiling produces the same quality at every length. The model documentation also says that minimal thinking is unsupported, so integrations should use Low, Medium, or High explicitly. 5

Gemini 3.8 Flash Cyber has a separate access boundary. Google says the model is available through the Fairwind Program to trusted defenders, government authorities, critical-infrastructure operators, and software maintainers. Google reports 47.2% on CWE-Bench pass@1, close to a leading frontier model's 47.8% at a lower cost, and says Chrome Security found 2.6 times more correct patches than larger commercial models. Those vendor and partner results support a defensive patching evaluation; they do not establish open access to offensive cyber work. 4

Muse Spark 1.3: the agent that asks for help

Meta released Muse Spark 1.3 on September 2 in Muse Code and the Meta Model API. Meta frames the release around long-horizon agent work rather than a new context-window or price headline. Muse Spark 1.3 can build context from messy sources, revise its plan when gaps appear, ask clarifying questions, ask the user for help when it is stuck, and confirm before consequential actions. 6

Meta says the model uses about 20% fewer tool calls and 25% fewer tokens on its coding evaluations than Muse Spark 1.2. The launch post gives a scorecard comparing Muse Spark 1.3 with Muse Spark 1.2, GPT-5.6 Sol, and Opus 5, but the public text does not state a price or context-window limit. That leaves an important measurement for the reader's own test: whether fewer calls still produce the same accepted result when the task contains conflicting files, a failed tool, and a late instruction change. 6

Muse Spark 1.3 is worth testing when an existing agent fails through coordination rather than raw knowledge. Give it a multi-step coding or operations task with an ambiguous requirement. Count unnecessary tool calls, clarification requests, irreversible actions, recovery after a tool error, final correctness, and total tokens. Meta's description points toward a model that spends more effort deciding when to act; the test should measure whether that caution saves rework or simply adds turns.

What to test first

Choose GPT-6 Astra when a long-running coding or computer-use workflow spends enough on tokens and retries for Astra's lower token count to matter. Compare total cost per accepted task, not the input price alone. 7
Choose Claude Fable 5.1 when code, research, or document work reuses large prompts. Measure cache hits, effort level, intervention rate, and cost per completed task. Apply for Mythos only when the restricted cyber or life-science route fits the organization. 3
Choose Gemini 3.8 Flash when one low-cost API must handle mixed media, long inputs, search, and tools. Test quality as context grows, because a large limit is an input boundary rather than a promise of uniform accuracy. 5
Choose Muse Spark 1.3 when the bottleneck is an agent's planning, recovery, and user coordination. Test it against the failure modes that create human cleanup work. 6

This week produced four evaluation paths rather than one universal winner. Astra makes token efficiency expensive but measurable. Fable 5.1 lowers the cost of premium agent work through cache reads and effort controls. Gemini 3.8 Flash makes long, multimodal, tool-using workflows cheap to start. Muse Spark 1.3 puts coordination behavior at the center of the test. A small run on the reader's own tasks will settle which trade-off matters before a migration does.