Anthropic Holds All Four Top Intelligence Slots on Artificial Analysis as Opus 5.5 and Sonnet 5.5 Sweep October
Artificial Analysis updated its model leaderboard this week, and the intelligence tier now shows a clean sweep. Claude Opus 5.5 (max with fallback) holds the top intelligence slot. Claude Sonnet 5.5 (max with fallback) sits second. Claude Opus 5.5 (xhigh with fallback) is third. Claude Opus 5.5 (high with fallback) rounds out the top four. All four belong to Anthropic.
That is not a rounding error or a methodology quirk. Artificial Analysis tests models independently using a ten-evaluation composite that includes AA-Briefcase, GDPval-AA, AutomationBench-AA, Terminal-Bench 4.0, Humanity’s Last Exam, SciCode, and four additional task-specific benchmarks. The rankings are not self-reported.
The Consolidation Timeline
This is not how it looked six months ago. In April 2026, Artificial Analysis showed a three-way tie at the top of the intelligence index between OpenAI, Anthropic, and Google. In June, Fable 5 pushed Anthropic to the top two positions. By August, Claude Opus 5 and Fable 5 held positions one and two.
The sweep to four positions happened with the late-September 2026 launches. Claude Opus 5.5 became available on 22 September at $4 per million input tokens and $20 per million output tokens. Claude Sonnet 5.5 followed on 28 September at the same price as its predecessor Sonnet 5. Within days of both launches, all four spots in the intelligence top tier shifted to Anthropic configurations.
The Sonnet Detail
The second position is the more striking result. Claude Sonnet 5.5 (max with fallback) placing above Claude Opus 5.5 (xhigh with fallback) means effort level is now a larger variable than model tier for a significant band of tasks. Running a smaller model at maximum inference effort outperforms running a larger model at elevated-but-not-maximum effort. Artificial Analysis documents this effect cleanly because it tests specific model configurations rather than just model versions.
For buyers, the implication is direct. If Sonnet 5.5 at max effort beats Opus 5.5 at xhigh effort on intelligence benchmarks, the cost calculus for many workloads shifts toward effort-scaling over model-scaling. Sonnet 5.5 is priced at $10 per million output tokens. Opus 5.5 is $20.
The Speed Tier Is a Different Market
The output speed rankings look nothing like the intelligence rankings. Celeris-1 is first. Mercury 2 is second. Mercury 2.5 is third. Gemini 2.5 Flash-Lite is fourth. Anthropic does not appear in this list.
Celeris-1 is a diffusion-based language model with verified output throughput above 1,600 tokens per second on commodity GPU infrastructure. Mercury 2 from Inception Labs uses a similar non-autoregressive architecture. Gemini 2.5 Flash-Lite is Google’s inference-optimized variant, not a frontier reasoning model.
The separation is structural. Models optimized for intelligence benchmarks require autoregressive generation, extended thinking budgets, and large parameter counts. Models optimized for throughput use parallel decoding or speculative approaches that sacrifice the sequential reasoning depth that drives benchmark performance. The same architecture cannot lead both lists.
What This Means for Procurement
Enterprise AI buyers in late 2026 are effectively choosing between two distinct product categories when they evaluate frontier models.
The intelligence buyer — legal, medical, research, code synthesis — follows the Anthropic rack. Every configuration they might reach for in the top tier is a Claude product. The competitive pressure from OpenAI and Google in this band, which was tangible as recently as six months ago, has contracted to single-digit gaps or below the measurement threshold.
The throughput buyer — translation, summarisation, classification at scale, real-time inference — follows a different rack. Celeris-1 and Mercury 2 are not competing for intelligence benchmark positions. They are competing on tokens per second per dollar.
Whether the two markets converge again depends on whether autoregressive reasoning models find credible speed improvements without architectural compromise. Celeris-1’s October position — still first on output speed, about two and a half months after its July launch — suggests the diffusion model cohort has found a durable market position, not a temporary one.
Anthropic holds the intelligence tier for now. The question for the next quarter is whether that sweep extends further, or whether OpenAI and Google bring Opus 5.5-level performance at Sonnet 5.5 prices.