Claude Opus 5.5 leads as rivals compete on cost

Claude Opus 5.5 leads the Artificial Analysis Intelligence Index results checked on 3 October 2026, scoring 58 and standing out for coding and long-form writing. GPT-6.1 Sol scores 52, close to GPT-6 Astra at 53, while costing $0.72 per task against Astra’s $3.26. Opus costs $5.98 per task in the same evaluation.

Token prices alone can mislead. Claude Sonnet 5.5 scores 56 and lists at half Opus’s token price, but its maximum-effort evaluation costs $7.67 per task. Claude Haiku 5.5 offers a budget option for classification, extraction and routing, scoring 43 at $0.21 per task. Its verbose output and prompt-length pricing require attention when estimating production costs.

Google opened Gemini 4 Argon to paid API customers on 13 October. Its index score of 53 was measured at high effort, unlike the maximum-effort Claude results. Introductory input and output pricing is $2 / $10 per million tokens through 30 November, then $4 / $20. Long-output workloads are a key reason to test it.

Self-hosting requirements narrow the field to downloadable models such as DeepSeek V4 and GLM-5.3. Mistral Large 4 remains a hosted preview with weights promised and no licence named. Model selection should prioritize data-handling requirements, actual task quality, latency and measured workload costs rather than leaderboard position alone.