📊 Verified Benchmark Scorecard
Every cell is a dated third-party snapshot on CompareLLM’s daily 06:00 UTC ingest. CompareLLM does not run Arena battles or SWE-bench in-house; see methodology. Chat Elo and Coding Elo: LMArena, 16 Aug 2026 (GPT-5: 17 Aug 2026). SWE-bench and LiveBench: 16 Aug 2026 (GPT-5: 17 Aug 2026). GPQA and speed: seed-bootstrap, 16 Aug 2026 (GPT-5: 1 Aug 2026). Prices: OpenRouter, 17 Aug 2026, except Gemini 3.6 Pro at $1.25/$5 (seed-bootstrap, 16 Aug 2026).
| Model | Chat Elo | Coding Elo | SWE-bench | LiveBench | GPQA | Speed | In / Out |
|---|---|---|---|---|---|---|---|
| Claude Opus 5 | 1,624 | 1,618 | 79.2% | 74.8% | 87.1% | 72 tok/s | $10 / $50 |
| Claude Opus 4.8 | 1,598 | 1,604 | 78.4% | 73.6% | 86.2% | 66 tok/s | $5 / $25 |
| GPT-5.6 Sol | 1,608 | 1,602 | 77.6% | 73.9% | 86.8% | 84 tok/s | $5 / $30 |
| Grok 4.6 | 1,592 | 1,568 | 69.1% | 71.4% | 84.6% | 118 tok/s | $2 / $6 |
| Gemini 3.6 Pro | 1,570 | 1,566 | 74.6% | 71.8% | 84.4% | 102 tok/s | $1.25 / $5 |
| GPT-5 | 1,558 | 1,540 | 68.4% | 69.8% | 85.2% | 78 tok/s | $1.25 / $10 |
Same-snapshot context: Opus 5 and Opus 4.8 at 1M, Sol at 1.1M, Gemini 3.6 Pro at 2M, Grok 4.6 at 500k, GPT-5 at 400k. TTFT: Opus 5 310 ms, Opus 4.8 330 ms, Sol 270 ms, Grok 4.6 195 ms, Gemini 3.6 Pro 205 ms, GPT-5 290 ms.
📈 Benchmark Analysis & Movement Drivers
The Claude Opus 5 vs Claude Opus 4.8 gap is a same-family increment, not a new capability class. Preference Elo is +26 (1,624 vs 1,598). Coding Elo is +14 (1,618 vs 1,604). SWE-bench is +0.8 points (79.2% vs 78.4%). LiveBench is +1.2 (74.8% vs 73.6%). GPQA is +0.9 (87.1% vs 86.2%). Decode speed rose from 66 to 72 tok/s; TTFT improved from 330 ms to 310 ms.
Versus OpenAI’s current flagship, Claude Opus 5 vs GPT-5.6 Sol is cleaner on quality and worse on the stopwatch: +16 Chat Elo, +16 Coding Elo, +1.6 SWE-bench (79.2% vs 77.6%), +0.9 LiveBench, +0.3 GPQA. Sol remains faster (84 vs 72 tok/s, 270 vs 310 ms) with a 1.1M window against Opus 5’s 1M. In the 72-model pool the published top three on 16 Aug 2026 are Claude Opus 5 at 1,624, Claude Fable 5 at 1,616, and GPT-5.6 Sol at 1,608.
Price—not Elo—is the cell in motion. Opus 4.8 is still $5/$25 on OpenRouter. Opus 5’s official Anthropic API is also $5/$25 with a 1M window (24 Jul 2026), but the 17 Aug 2026 OpenRouter snapshot still posts $10/$50, a clean 2× versus the prior Opus generation. A reported router input cut from $10 to $5 would erase the prompt premium versus Opus 4.8 and Sol. Until the next 06:00 UTC ingest confirms that cell, $10/$50 is the last verified router quote and $5/$25 is the official API quote.
Cost and throughput foils are unchanged. Grok 4.6 ($2/$6, 118 tok/s) sits at 1,592 Elo and 69.1% SWE-bench. Gemini 3.6 Pro ($1.25/$5, 2M context) is 1,570 Elo and 74.6% SWE-bench. Prior-flagship GPT-5 at $1.25/$10 is 1,558 Elo and 68.4% SWE-bench. None of those rows threaten the 1,624 ceiling.
🛠️ Impact on Real-World Workflows
Interactive chat. A 1,624 Arena score is a pairwise win-rate, not a latency SLA. Opus 5 is the default when preference margin beats 310 ms TTFT and 72 tok/s. Latency-bound sessions still favor Sol (270 ms / 84 tok/s) or Grok 4.6 (195 ms / 118 tok/s).
Coding assistants. The 1,618 Coding Elo and 79.2% SWE-bench keep Opus 5 at the top of the Best Coding LLM shortlist. The lift over Opus 4.8 is only +14 Coding Elo and +0.8 SWE-bench—visible on long-horizon patches, rarely on single-file completion. Teams already on Opus 4.8 at $5/$25 should fail an internal eval before paying a 2× OpenRouter premium.
Automated agents. Both Opus generations expose 1M context; Sol offers 1.1M and Gemini 3.6 Pro offers 2M. Agent bills are input-heavy. At the last verified $10 input cell, a 200k-token trace costs 2× Opus 4.8 or Sol on the prompt side. If the $10 → $5 input print lands, that penalty disappears and the residual tax is output ($50 vs $25 on the last OpenRouter card). Cost-first loops should stay on the cheap coding stack and Best Cheap LLM routes—Grok 4.6, Gemini 3.6 Pro, or GPT-5—until an eval gate actually requires 79.2% SWE-bench.
🎯 Platform Verdict
In the mid-2026 flagship band, Claude Opus 5 is the preference and SWE-bench leader, not the cost or latency leader. Route it when a +26 Elo / +0.8 SWE-bench increment over Opus 4.8, or a +16 Elo / +1.6 SWE-bench increment over Sol, changes a ship/no-ship gate. Otherwise keep Opus 4.8 on the official $5/$25 card, or drop a class. Re-read the OpenRouter input cell after the next 06:00 UTC ingest before locking Q3 traffic.
❓ Frequently Asked Questions
Is Claude Opus 5 better than Claude Opus 4.8 on benchmarks? Yes, on every dated cell here. As of 16 Aug 2026: 1,624 vs 1,598 Chat Elo, 1,618 vs 1,604 Coding Elo, 79.2% vs 78.4% SWE-bench, 74.8% vs 73.6% LiveBench, 87.1% vs 86.2% GPQA. The deltas are incremental. Full grid: Claude Opus 5 vs Claude Opus 4.8.
What is the Claude Opus 5 price in 2026? Official Anthropic API: $5 input / $25 output per 1M tokens, 1M context (24 Jul 2026). OpenRouter snapshot, 17 Aug 2026: $10 / $50. A reported router input move from $10 → $5 is the cell to confirm on the next daily ingest. Opus 4.8 remains $5 / $25 on that router.
Does Claude Opus 5 beat GPT-5.6 Sol? On the 16 Aug 2026 snapshots, yes for preference and coding: 1,624 vs 1,608 Elo, 1,618 vs 1,602 Coding Elo, 79.2% vs 77.6% SWE-bench. Sol is faster (84 vs 72 tok/s, 270 vs 310 ms) and slightly longer-context (1.1M vs 1M). See Claude Opus 5 vs GPT-5.6 Sol.
Should I switch coding traffic to Opus 5 today? Only if Opus 4.8 or Sol fails your SWE-bench-style eval. The SWE-bench gap is 0.8 points versus Opus 4.8 and 1.6 versus Sol. At the last verified OpenRouter quote you pay 2× input and 2× output versus Opus 4.8 for that margin.