September AI Race: Five Labs Enter the Field
Claude Fable 5.1 set a new benchmark as Gemini and Muse closed in. With GPT Astra and Grok 4.7 on deck, September has become a five-lab frontier race.
Table of Contents
Chinese frontier labs led the industry's focus throughout the summer of 2026. DeepSeek, Z.ai, and Moonshot AI narrowed the performance gap with faster, lower-cost releases—a trend examined in AIScroll's account of the summer shift in China. Leading American labs continued shipping updates, but a major reshuffling at the top of the leaderboard remained on pause.
September broke that lull immediately. Anthropic released Claude Fable 5.1 on September 1, followed a day later by Google's Gemini 3.8 Flash and Meta's Muse Spark 1.3. With OpenAI's GPT Astra and xAI's Grok 4.7 nearing release, the race has opened into a five-way contest between established incumbents and fast-moving challengers.
Fable Model Performance Leads at 66
The frontier model leaderboard has an opening leader: Claude Fable 5.1. Its maximum-reasoning configuration scored an all-time high of 66 on the Artificial Analysis Intelligence Index, four points above Fable 5 and surpassing Claude Opus 5 max at 63. The model also posted 59.1% on HLE, 62.0% on SciCode, and 91.4% on Terminal-Bench v2.1, leading across general reasoning, scientific problem-solving, and coding.
That top-tier performance comes at a steep price. Fable 5.1 max averaged $3.69 per Intelligence Index task, well above competing models. Anthropic defended the premium with a 1-million-token context window, enhanced long-horizon agent execution, and a 75% price reduction for prompt caching. Even with roughly 4% of evaluation output routed to other Claude models for safety, Fable secured the early September lead.
Yet 66 points did not close the contest. Challengers trailing Fable traded a few index points for sharply lower operational costs and faster inference speeds.
| Model and setting | Intelligence Index | Cost per task | Status |
|---|---|---|---|
| Claude Fable 5.1 max | 66 | $3.69 | Overall leader |
| Muse Spark 1.3 max | 62 | Undisclosed | Limited preview |
| Muse Spark 1.3 xhigh | 61 | $0.55 | Generally available |
| Gemini 3.8 Flash high | 59 | $0.58 | Generally available |
| Grok 4.6 high | 61 | $0.94 | Baseline for Grok 4.7 |
Gemini Rebounds: 3.8 Flash Rejoins the Frontier Hunt
Google spearheaded that pursuit. Delays surrounding the next Pro model and the launch of the cost-oriented 3.7 Flash had fueled sentiment that Google was falling behind at the frontier. Gemini 3.8 Flash countered that narrative just three weeks after its predecessor debuted.
Gemini 3.8 Flash high recorded 59 on the Intelligence Index, three points above 3.7 Flash. The Google Gemini benchmarks also logged 73.7% on DeepSWE v1.1, 89.4% on Terminal-Bench 2.1, and 54.9% on HLE-Verified. Generating roughly 300 tokens per second at $0.58 per task, the model pairs high throughput with aggressive cost efficiency.
Taken together, the Google Gemini benchmarks still trail Fable's 66 points, but the comeback carries significance beyond one leaderboard position. By shipping three Flash iterations in six weeks, Google has turned what was once an efficiency-first tier into a genuine frontier contender.
| Model | Benchmark | Score | Context |
|---|---|---|---|
| Claude Fable 5.1 | Terminal-Bench v2.1 | 91.4% | Independent Artificial Analysis run |
| Gemini 3.8 Flash | DeepSWE v1.1 / Terminal-Bench 2.1 | 73.7% / 89.4% | Google-reported |
| Muse Spark 1.3 | DeepSWE v1.1 / Terminal-Bench 2.1 | 75.4% / 88.8% | Meta-reported |
| Claude Fable 5.1 | Terminal-Bench 4.0 | 55.8% | Not comparable with v2.1 |
Meta's Surprise Surge Reaches the Frontier
Google was not alone in staging a dramatic move. Meta's Muse Spark 1.3 max reached 62 on the Intelligence Index, trailing only Fable 5.1 and Opus 5 tiers. Its widely accessible xhigh configuration scored 61, matching GPT-5.6 Sol max and Grok 4.6 high. That leap propelled a provider once viewed as a late entrant directly into the frontier echelon.
Granular benchmarks backed that composite score. Muse Spark 1.3 posted 75.4% on DeepSWE v1.1, edging Gemini, and scored 88.8% on Terminal-Bench 2.1. At $0.55 per task, the xhigh tier is the lowest-cost model scoring 59 or higher on the index. Meta also cut tool calls by roughly 20% and token consumption by 25% compared with Muse Spark 1.2.
The impact extends well beyond claiming third place. Meta has committed to releasing Muse Spark 1.3 as open weights. If public weights match these benchmark results, Meta will exert pricing and performance pressure on commercial APIs and open-source models alike. Rather than following the leaders, Meta is directly rewriting the rules of the leaderboard.
GPT Astra: 'Critical' Rating Fuels GPT-6 Expectations
The three released models represent only the opening chapter of September's contest. The largest wildcard is OpenAI's GPT Astra, previewed as an impending release in an official capability update. Although broad reasoning and coding benchmark totals remain under wraps, early technical disclosures point to a distinct generational shift.
Astra is the first OpenAI model to earn a Critical designation in cybersecurity capability. It achieved a 100% success rate on ExploitBench across 41 known vulnerabilities. Tested against 20 recent V8 flaws, Astra delivered higher arbitrary code-execution rates than GPT-5.6 Sol with fewer tokens, while unearthing two zero-day vulnerabilities in the process.
OpenAI has not confirmed GPT-6.0 as the model's official release name. Nevertheless, the elevated risk rating and performance gains over GPT-5.6 Sol have fueled expectations that Astra represents the next major generational leap. If public benchmarks validate those internal metrics upon release, the 66-point ceiling established early this month could quickly fall.
Grok 4.7 Launch Completes the Five-Way Contest
Astra is not the only challenger lined up for the second wave. On September 2, Elon Musk said the Grok 4.7 launch would come in ten days, pointing to a debut around September 12. Because the preceding Grok 4.6 high had already scored 61 on the Intelligence Index to tie GPT-5.6 Sol max, expectations for the new version's performance gain run high.
Official model cards and granular benchmark tables have yet to appear. If xAI meets Musk's timetable, the Grok 4.7 launch will open a second ranking contest immediately after the first-week clash among Fable, Gemini, and Muse.
The summer surge from Chinese labs pushed American Big Tech to accelerate. The frontier competition now extends beyond Anthropic, OpenAI, and Google to include Meta and xAI. As former latecomers pull level with the leading pack, the market has entered its most volatile stretch of the year.
Fable 5.1, Gemini 3.8 Flash, and Muse Spark 1.3 arrived in rapid sequence during the first week of September, with GPT Astra and Grok 4.7 next. Rather than cementing a fixed frontier model leaderboard, September has started a release cycle in which the top spot could shift several times before month's end. The AI race has entered a new phase.
