Three Frontier Models in 72 Hours: What Changed in September 2026
AI & Infrastructure
Between September 1 and September 3, three labs shipped new flagship models. Anthropic released Claude Fable 5.1 and Mythos 5.1. Google released Gemini 3.8 Flash and 3.8 Flash Cyber. OpenAI released GPT-6 Astra. Most of the coverage went to benchmark scores. The change with more practical weight is how access to the strongest capabilities is now being distributed.
⚡ TL;DR
| Release | Date | What matters |
|---|---|---|
| Claude Fable 5.1 + Mythos 5.1 | Sep 1 | Input and output prices unchanged at $10/$50. Cache reads fell 75%, which changes agent economics. |
| Gemini 3.8 Flash + Flash Cyber | Sep 2 | Same price per token, roughly 40% higher real cost per task. Intro pricing ends Dec 31. |
| GPT-6 Astra | Sep 3 | Large vendor benchmark claims, contested harnesses, first OpenAI model above the Critical cyber threshold. |
All three launches follow the same shape: a public model plus a restricted cybersecurity tier running on the same weights.
🟠 Anthropic: Same Rate Card, Different Bill
On September 1 Anthropic shipped two models on identical weights.
Claude Fable 5.1 is generally available with production safeguards. Claude Mythos 5.1 has the same capabilities with relaxed safeguards and is restricted to Project Glasswing participants, meaning vetted cyberdefense and life sciences organizations.
Fable 5.1 specs: 1M token context, 128K max output, adaptive thinking always on, default effort set to high, knowledge cutoff June 2026.
Input and output pricing did not change. It stays at $10 and $50 per million tokens. The change that affects most production workloads is in a pricing footnote:
Cache reads fell from $1.00 to $0.25 per million tokens, a 4x cut.
Anthropic estimates around 25% savings on typical workloads and up to 45% on heavy agentic and coding work. Those figures come from Anthropic's own August usage measurement, so treat them as vendor estimates tied to a specific workload mix. The direction holds regardless. For a long-running agent that carries the same system prompt and history across hundreds of steps, cache reads make up most of the bill.
Current lineup
| Model | Context | Max output | Input / Output | Cutoff |
|---|---|---|---|---|
| Claude Fable 5.1 | 1M | 128K | $10 / $50 | Jun 2026 |
| Claude Opus 5 | 1M | 128K | $5 / $25 | May 2026 |
| Claude Sonnet 5 | 1M | 128K | $2 / $10 | Jan 2026 |
| Claude Haiku 4.5 | 200K | 64K | $1 / $5 | Feb 2025 |
One detail most reviews skip: Anthropic recommends starting with Opus 5 and moving to Fable 5.1 only for demanding reasoning and long-horizon agentic work, or when evals on Opus 5 at higher effort still fall short. Fable costs twice as much and runs noticeably slower.
Published gains
Terminal-Bench 4.0 42.0% -> 55.8% Terminal-Bench-Science 0.1 24.7% -> 52.6%
Breaking changes
Three items in the migration guide will return 400 errors on existing agent code:
- Forced tool use
- Model-bound thinking blocks
- History editing
Read the migration guide before bumping the model string.
🔵 Google: Third Flash Release in Six Weeks
Google shipped Gemini 3.8 Flash on September 2, its fourth Flash release since May. Introductory pricing holds at $0.75 and $3.75 per million tokens, with a 1M context window and three thinking levels.
Terminal-Bench 2.1 81.6% -> 90.8% HLE-Verified 54.9%
On several rows a Flash-tier model now lands within a couple of points of Claude Opus 5, which is unusual for that price bracket.
Two caveats change how those numbers read.
First, the model card treats 3.8 Flash as further training on 3.7 Flash rather than a new base model. The architecture, training data, hardware and safety sections all defer to the previous card.
Second, price per token does not translate to price per task here. The model reasons longer and calls tools more often, so Artificial Analysis measured real costs roughly 40% higher than 3.7 Flash at identical token rates. Google says the same thing in its own guidance: stay on 3.7 Flash for high-volume, short-context, latency-sensitive workloads.
Introductory pricing runs through December 31, 2026. On January 1 both rates double, to $1.50 and $7.50.
The next-generation flagship Gemini Pro, expected in June, has still not shipped.
🟢 OpenAI: Big Claims, Contested Numbers
OpenAI released GPT-6 Astra on September 3 and described it as its most intelligent and most aligned model. The positioning is computer use rather than conversation: browsing, software engineering, multi-step professional work. It is the company's largest training run to date at over 100,000 GPUs, and the first release where earlier OpenAI models supervised training of the new one.
The vendor numbers are large.
| Benchmark | GPT-5.6 Sol | GPT-6 Astra |
|---|---|---|
| FrontierMath Tier 4 | 83.0% | ~98% |
| Terminal-Bench 4.0 | 37.3% | 57.9% |
| Science workflow | 22.4% | 64.6% |
| ExploitBench | 78.5% | 100% |
The harness problem
The headline 99.9% on ARC-AGI-3 was produced on a provider adapter harness that preserves private reasoning state and compacts long conversations. ARC Prize measured 62.7% on its standard harness.
Independent testing shows a flatter result. Artificial Analysis put Astra at 61.2 against 60.9 for its predecessor on the broad intelligence index, which is inside the noise. The coding agent index shows a wider gap at 67.0 against 65.1. Fortune also found that several published figures were revised after launch.
Comparing releases using vendor tables does not work when the harness accounts for more of the difference than the model does.
The cyber threshold
Astra is the first OpenAI model to cross the Critical threshold in the company's Preparedness Framework. It scored 100% on ExploitBench against 70% for Fable 5.1, and found two previously unknown zero-days during evaluation. That is the reason for the staged rollout, with advanced cyber capability gated behind the Daybreak program.
Reduced monitorability
The system card reports that following the model's chain of reasoning became harder, particularly when it attempts evasion. Capability went up while one of the primary signals researchers use to detect misalignment got weaker. That trade-off does not appear in any benchmark table and deserves more attention than it has received.
🧭 Intelligence and Permission Are Splitting
All three launches pair a general model with a restricted cyber tier built on the same weights.
| Lab | Public | Restricted | Program |
|---|---|---|---|
| Anthropic | Fable 5.1 | Mythos 5.1 | Project Glasswing |
| 3.8 Flash | 3.8 Flash Cyber | Fairwind (650+ orgs, including CrowdStrike, Datadog, Palo Alto Networks, Snowflake) | |
| OpenAI | Astra | Advanced cyber capability | Daybreak |
Vulnerability discovery and exploitation now sit at roughly the same level across all three labs. What differs is the access regime around them. Competition has expanded from raw capability to which governance model regulators and enterprise buyers will accept.
For anyone building automation on contract, this has a direct consequence. Once a project touches pentesting, malware analysis or reverse engineering, the blocker is program access rather than model quality, and applications come from organizations rather than individual developers.
🛠️ What To Do This Week
1. Recalculate cache costs, not token costs. With cache reads down 4x, the shape of an agent's bill has changed. If you have a long system prompt and a stable prefix, restructure so the cacheable segment is as long and as immutable as possible. This is currently the cheapest available cost reduction.
2. Do not default to the top model. Opus 5 costs half of Fable 5.1 and is Anthropic's recommended starting point. Upgrade when your own evals at higher effort hit a ceiling.
3. Build your own eval set. The ARC-AGI harness discrepancy and the post-launch revisions make the case on their own. Twenty real tasks from your domain will tell you more than any leaderboard.
4. Measure cost per task. With adjustable thinking levels, the rate card says little about the cost of completed work. Test at the effort level you plan to run in production.
5. Check breaking changes before upgrading. Fable 5.1 changes forced tool use, thinking blocks and history editing.
6. Budget for January. Gemini Flash introductory pricing ends December 31, 2026.
🎯 Summary
The first week of September delivered smaller quality gains than the headlines suggested and a larger structural change than the benchmarks show.
Two things are new. The cost of long-running agents is now driven by cache reads rather than headline token rates. And all three frontier labs decided at the same time that part of their capability will not be sold from a public price list.
The second one says more about where this industry is going than any table of percentages.
📚 Sources
- Claude Fable 5.1, Anthropic Platform Docs
- Introducing Gemini 3.8 Flash and 3.8 Flash Cyber, Google
- OpenAI launches Astra, TechCrunch
- GPT-6 Astra Benchmarks Explained, Vellum
- GPT-6 Astra: Benchmarks, Cyber Risks, and Market Impact, Futurum
- Google shipped four Gemini Flash models in 106 days, Fortune
- Gemini 3.8 Flash review 2026, eesel AI
- OpenAI's new model aces the benchmarks, The Next Web
Published on esso.dev. I help teams architect AI agent infrastructure that survives the next model release. Get in touch
Comments
Loading comments...