Breaking

Anthropic, Google and Meta Release New Coding Models in 48 Hours

September 3, 2026 at 13:01 EDT

Foundation Models
Software Dev & Coding
AI Agents

Anthropic, Google, and Meta released new coding-focused AI models within 48 hours, each pitching stronger long-horizon software work as the labs race to own agentic development. The cluster of launches, spanning September 1 and 2, 2026, puts frontier coding, cheaper tokens, and restricted cybersecurity variants at the center of the commercial fight.

September 2026 · Frontier models

Three labs, one coding race

Anthropic, Google, and Meta released new agentic coding models within 48 hours, each claiming stronger long-horizon software work at a lower effective cost.

Release flow

1 · Sep 1

Anthropic

Claude Fable 5.1 + Mythos 5.1

→

2 · Sep 2

Google

Gemini 3.8 Flash + Flash Cyber

→

3 · Sep 2

Meta

Muse Spark 1.3

52.6%

Fable 5.1 on Terminal-Bench-Science 0.1

75%

Cut in Fable cache-read price, to $0.25 / 1M

3rd

Gemini Flash release in six weeks

-25%

Muse Spark 1.3 tokens vs 1.2 on coding jobs

Terminal-Bench-Science 0.1

Same unit, drawn to scale. Each block is about 5 percentage points. Fable 5.1 more than doubles Fable 5.

22.4%

GPT-5.6 Sol

24.7%

Fable 5

29.0%

Opus 5

52.6%

Fable 5.1

Cache reads: 4× cheaper

Fable 5.1 keeps the $10 / $50 list price but cuts cache reads from $1.00 to $0.25 per million tokens.

$1.00

Fable 5 cache

$0.25

Fable 5.1 cache

Takeaway: typical Fable workloads fall about 25%; highly agentic jobs can fall about 45%.

What users like

Cognition is moving Devin’s Opus 5 traffic to Fable 5.1. Jane Street and Millennium report harder coding and debugging wins. Muse Spark 1.3 and Gemini 3.8 Flash undercut Fable on list price.

What still bites

Google says 3.8 Flash may burn more tokens at high effort. Meta’s max reasoning is still gated. Mythos 5.1 and Flash Cyber stay in trusted programs. Fable-class models hit a same-day error spike on September 3.

Cyber twins stay behind the gate. The strongest vulnerability-finding variants are not general-availability products.

Mythos 5.1 — same weights as Fable 5.1; U.S. trusted cyber and life-sciences access only.
Gemini 3.8 Flash Cyber — Fairwind Program for trusted defenders; 86.2% on CyberGym.
Fable 5.1 — may find software bugs, but exploit work still routes to Opus models.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.