OpenAI Ecosystem September 2 Briefing: Anthropic Launches Claude Fable 5.1 / Mythos 5.1, Benchmarks Include GPT-5.6 Sol

The main item in this September 2 briefing is competitor news: Anthropic announced Claude Fable 5.1 and Claude Mythos 5.1 on 2026-09-01, and listed GPT-5.6 Sol as a direct comparison target in its official benchmark table. This briefing relays only what Anthropic's official announcement supports -- model structure, pricing, data retention, access programs, and benchmark numbers -- and does not speculate about "what it means for OpenAI"; there is no citable new OpenAI announcement at publish time, so none is claimed. Our weekly briefings have long tracked competitor signals under their competitor-boundary coverage; this is the first standalone item in that line.

1. The One-Paragraph Version

Anthropic shipped Claude Fable 5.1 (generally available) and Claude Mythos 5.1 (trusted-access only) -- two safeguard tiers of the same model. Unit pricing holds at $10/$50 per MTok, but cache reads drop to $0.25/MTok (down 75% by the official count); measured over four weeks of real August 2026 usage, typical workloads save about 25% and highly agentic workloads up to about 45%. And the official benchmark table puts GPT-5.6 Sol in direct comparison.

2. Model Structure: One Model, Two Safeguard Tiers

The structural decision at the heart of the announcement: Fable 5.1 and Mythos 5.1 are not two model generations but two safeguard levels of one model.

The official wording for Mythos 5.1: safeguards designed to support cybersecurity and life-sciences work, with Anthropic coordinating with the US government to expand access to domestic and international partners and opening scientist enrollment soon.

3. Pricing: Same Unit Price, Cache Reads Cut 75%

The most bill-relevant part of this launch:

• Unit price unchanged: $10 per MTok (input), $50 per MTok (output) -- same as Fable 5.
• Cache reads drop: to $0.25 per MTok, officially described as 75% less. Cache reads cover reused input context the model has already processed and stored.
• Bill impact (official measurement): an index of four weeks of real August 2026 usage at default effort (Fable 5 = 100) puts typical workloads at about 75 (~25% less) and cache-heavy, tool-heavy highly agentic workloads at about 55 (~45% less). The official wording carries estimated.

For context from our world: GPT-5.6 Sol currently sits at a promotional $4/$20 per MTok (from 2026-08-21, committed at least through 2026-11-21). The two vendors' billing structures differ -- promotional windows, cache pricing -- so this briefing draws no converted conclusion; plug in your own workload structure.

4. Data Retention and Safety: EFS and Fewer False Positives

• EFS (Enterprise Frontier Safeguards): data is stored in cloud infrastructure controlled entirely by the customer -- not Anthropic -- with privacy the announcement describes as equivalent to a zero data retention policy, while maintaining state-of-the-art protection against adversarial use. It reaches enterprise customers in phases beginning later this fall; until EFS is available, eligible customers can use Fable 5.1 with zero data retention.
• Safeguard improvements: cybersecurity false positives (benign content flagged) are down 60%. The capability boundary shifts with them: Fable 5.1 can be used to discover software vulnerabilities -- though not to develop exploits for them.
• Life sciences: an access program developed in partnership with the US government opens Claude Mythos 5.1's advanced biology capabilities, with scientist enrollment expected to open soon.

5. The Official Benchmark Table: How to Read the GPT-5.6 Sol Column

The announcement's comparison table puts GPT-5.6 Sol alongside Fable 5.1, Fable 5, and Opus 5. The Sol column, relayed as printed (each benchmark under its officially noted setup):

Read the table with the four caveats the announcement itself attaches (expanded in our companion piece, How to Read the Benchmark Charts in a Model Launch: Effort Curves, Log Cost Axes, and Safeguard Interventions):

1. Safeguard interventions score zero: Fable 5.1 ran with production safeguards; tasks where they intervened scored zero -- on OSWorld 2.0 for both Fable tiers and on AutomationBench for Fable 5 -- which Anthropic says likely reduces those scores.
2. Two benchmarks have no Sol entry: OSWorld 2.0 and HLE show dashes for Sol -- a missing cell is not a loss, just an unreported score.
3. In-house setup vs public leaderboard: on Terminal-Bench-Science 0.1, the public leaderboard (3 trials/task, Claude Code harness) reports Opus 5 at 30.0% and Fable 5 at 21.4%, while Anthropic's own setup reproduces 29.0% and 24.7% -- within noise, per the footnote, but the two setups produce different numbers.
4. Error bars: Terminal-Bench-Science 0.1 carries a standard error of ±3.5-4.5 points per model.

6. Effort Tiers and Defaults

The announcement's cost curves break out by effort tier: low / med / high / xhigh / max, with official defaults -- High in Claude Code, Medium in Claude Cowork and on Claude.ai -- and the claim that at Low or Medium effort Fable 5.1 achieves results similar to or better than Fable 5 at much lower cost. This is the same design problem as GPT-5.6's continuous reasoning.effort: cross-vendor price comparisons must lock the same effort tier, or they do not hold.

7. The OpenAI Side This Week

This briefing has no new OpenAI announcement to link: this site's last successful API changelog check was on 2026-09-01, when the latest entry was 8-29's mTLS / X.509 general availability (see OpenAI Ecosystem Week 43 Briefing (2026-09-01): mTLS GA, Sora 2 Countdown, Shutdown Calendar); on 9-02 the developer docs domain rate-limited our fetches, so no re-check was possible and this piece makes no assertion about OpenAI's state beyond that check.

8. Next Steps

• The full method for reading launch benchmark charts (log cost axes, error bars, partial/strict scoring, missing-cell semantics): How to Read the Benchmark Charts in a Model Launch: Effort Curves, Log Cost Axes, and Safeguard Interventions.
• The GPT-5.6 family and this season's pricing: The complete guide to GPT models (2026-07): GPT-5.6 Sol, Terra, Luna.
• Last briefing: OpenAI Ecosystem Week 43 Briefing (2026-09-01): mTLS GA, Sora 2 Countdown, Shutdown Calendar.