Live AI agent hackathon · Monday, September 14, 2026
InventoryHero
1
Where we are
Two weeks that moved the whole board.
- Sep 1Anthropic ships Claude Fable 5.1 and Mythos 5.1.
- Sep 2Google ships Gemini 3.8 Flash.
- Sep 3OpenAI ships GPT-6 Astra. "A generational leap."
- Sep 8OpenAI publishes a checked proof on Navier-Stokes.
- Sep 10OpenAI ships GPT Image 2.5, Flare and Sunburst.
InventoryHero
2
Section 1 · OpenAI's two weeks
GPT-6 Astra: OpenAI calls it a generational leap.
Computer use
Drives a browser and a desktop.
State of the art at navigating computers and the web.
Cyber
First model rated Critical.
Finds unknown flaws on its own. That is why the release slipped.
Price
$10 in, $50 out.
Per million tokens. Identical list price to Fable 5.1.
The catch
"Recurrent depth" hides its reasoning.
Safety researchers cannot watch all of its thinking.
InventoryHero
3
Section 1 · OpenAI's two weeks
Astra helped close a 200 year old math problem.
Navier-Stokes: the equations for how fluids move. A $1M Millennium Prize question.
10,000
agents running at once
88 hrs
to the proof, on an unreleased model
17 hrs
for Astra to write it in Lean so a computer could check every step
2nd
Millennium Prize problem ever resolved
Terence Tao, as reported: "a carcass of raw meat on the table." True, but no why. OpenAI is not claiming the prize.
InventoryHero
4
Section 1 · OpenAI's two weeks
GPT Image 2.5 owns product and branding images.
| Rank | Model | Lab | Score |
|---|---|---|---|
| 1 | gpt-image-2.5-sunburst | OpenAI | 1419 |
| 2 | gpt-image-2 (medium) | OpenAI | 1389 |
| 3 | gpt-image-2.5-flare | OpenAI | 1388 |
| 4 | mai-image-2.6 | Microsoft AI | 1336 |
| 8 | gemini-3.1-flash-image (nano-banana-2) | 1269 |
Arena.ai Text-to-Image, Product, Branding & Commercial Design. Sep 7, 2026. 1.9M votes, 78 models. Sunburst and Flare preliminary.
Sunburst is the quality model. Flare is the fast one, half the latency of GPT Image 2.
InventoryHero
5
Section 2 · Anthropic's answer
Claude Fable 5.1: same brain, built for long jobs.
Shipped Sep 1
Coding, knowledge work, long runs.
Fixes root causes instead of taking shortcuts.
Price
$10 in, $50 out.
Unchanged from Fable 5. Cache reads dropped 75% to $0.25.
Real cost
25% to 45% cheaper.
Typical work, and the top end is for autonomous agent runs.
Mythos 5.1
Same model, fewer guardrails.
Gated to trusted programs. Fable is what you get in Claude Code.
InventoryHero
6
Section 3 · Head to head
On real agent work, Fable leads by a nose.
| Rank | Model | Net improvement | Confirmed success | Praise vs complaint |
|---|---|---|---|---|
| 1 | Claude Fable 5.1 (Max) | 13.90% | 23.70% | 30.47% |
| 2 | GPT-6 Astra (Max) | 11.90% | 19.54% | 34.41% |
| 3 | Claude Opus 5 (Max) | 11.09% | 12.92% | 19.03% |
Arena.ai Agent Arena, Overall. Sep 13, 2026. 1.69M sessions, 43 models.
Rank spreads overlap (1 to 4 vs 1 to 6). Call it a statistical tie at the top.
InventoryHero
7
Section 3 · Head to head
Tied on smarts. Astra is cheaper per task.
| Fable 5.1 | GPT-6 Astra | Opus 5 | Gemini 3.8 Flash | |
|---|---|---|---|---|
| Intelligence Index | 53 | 53 | 51 | 41 |
| Speed, tokens per second | 65 | 60 | 51 | 278 |
| Cost per task | $7.63 | $3.26 | $5.86 | $1.24 |
artificialanalysis.ai/models, Highlights. Viewed Sep 14, 2026. Max effort configurations; Fable 5.1 with fallback.
Same list price. Fable spends more tokens thinking.
InventoryHero
8
Honorable mention
Gemini 3.8 Flash is the value play.
41
Intelligence Index. Frontier level a few months ago.
278
tokens per second. Fastest on the chart.
$1.24
per task. A quarter of Astra, a sixth of Fable.
$0.75
in, $3.75 out per million, through Dec 31.
artificialanalysis.ai and blog.google, Sep 2026.
InventoryHero
9
Close
Pick the tool by the job.
Images, listings, one-shot tasks
GPT-6 Astra + GPT Image 2.5
Long agent runs and code
Claude Fable 5.1
Volume, speed, cheap
Gemini 3.8 Flash
Now, live. Replays and the next session:
inventoryhero.ai/ai-agents