On the morning of October 23, Claude Opus 5.5 slipped onto the Hacker News front page without much fanfare. 135 points, 43 comments — not exactly viral, but enough to spark a heated debate among developers.
The argument wasn't whether Opus 5.5 is strong (it is). It was about three more practical questions:
- How much better is it than Opus 4.5?
- Has the gap with GPT-5 and Gemini 2.5 Pro widened or narrowed?
- How much more expensive is it, and is the premium worth paying?
Pulling data from artificialanalysis.ai, here's a breakdown of what the numbers actually show.
What HN Developers Are Arguing About
The HN thread clustered around a few recurring topics:
- "Do Opus 4.5 users have any reason to upgrade?" — the top concern for enterprise customers
- "Coding benchmark gap vs GPT-5" — the favorite comparison in dev circles
- "Sonnet already does the job, so why bother with Opus?" — the classic cost-conscious objection
- "Is the Max-tier pricing an anchoring play?" — a rare thread of actual business analysis
The top-voted comment captured the mood: "This is a 5% improvement at 30% higher cost." That sentiment showed up more than once.
Number 1: Did Intelligence Actually Improve?
artificialanalysis's intelligence index is a composite score that weights multiple independent benchmarks (MMLU-Pro, GPQA, AIME, LiveCodeBench, etc.). Here's how Opus 5.5 stacks up against Opus 4.5, GPT-5, and Gemini 2.5 Pro:
| Model | AA Intelligence Index | vs Opus 5.5 |
|---|---|---|
| Claude Opus 5.5 | 73 | — |
| Claude Opus 4.5 | 71 | -2 |
| GPT-5 | 70 | -3 |
| Gemini 2.5 Pro | 68 | -5 |
The first number: 73. Two points above 4.5, three above GPT-5, five above Gemini.
Does this gap mean anything? Statistically, it's borderline significant. But in real workflows, a 2-point difference is small enough that better prompt engineering can erase it.
The verdict on intelligence: Opus 5.5 holds the top spot but didn't pull ahead. The real advantage lies in overall product capability, not raw intelligence scores.
Number 2: Throughput — The Underrated Metric
Anthropic advertises an output speed of "~80 tokens/s." What does real-world performance look like?
AA measured actual throughput across three scenarios — long context (128K), Tool Use, and Agent loops:
- 128K long-context inference: ~52 tokens/s average, 1.8s time-to-first-token
- Tool Use (10 consecutive turns): stable at 68-72 tokens/s
- Agent loop (single-step decision→action→observation): 47 tokens/s
The second number: 47 tokens/s. This is the real-world throughput in Agent scenarios — about 40% below the official spec.
Why does this matter? In 2026, people aren't just using Claude to "write a block of code" anymore. Claude Code, the Claude backend in Cursor, and various Agent frameworks (LangGraph, AutoGen) all hammer the API in repeated loops.
Opus 5.5 is roughly 18% faster than 4.5 in Agent scenarios, but the time-to-first-token barely changed (1.7s → 1.8s, actually slightly worse).
So the curve looks like this: long-context gets a modest speedup, Tool Use is unchanged, and time-to-first-token didn't improve. The speed gain is conditional at best.
Number 3: Pricing — 30% More Expensive or 30% Cheaper?
Here's the raw data (USD per million tokens):
| Model | Input | Output | Cache Read |
|---|---|---|---|
| Claude Opus 5.5 | $15 | $75 | $7.50 |
| Claude Opus 4.5 | $12 | $60 | $6 |
| Claude Sonnet 5 | $3 | $15 | $0.30 |
| GPT-5 | $10 | $30 | $2.50 |
| Gemini 2.5 Pro | $5 | $20 | $1 |
The third number: output went from $60 to $75, a 25% increase.
A few observations:
- vs Opus 4.5: 25% more expensive
- vs GPT-5: output is 2.5x the price
- vs Gemini 2.5 Pro: output is 3.75x the price
- vs Sonnet 5: 5x more expensive (though Opus was never meant for high-frequency calls)
One side note: the $7.50 Cache Read rate looks steep, but for long-context Agent workloads where cache hit rates exceed 60%, it can offset a meaningful chunk of cost.
Price-performance ratio (intelligence index ÷ output price):
| Model | Intelligence Index | Output Price | Price-Performance |
|---|---|---|---|
| Gemini 2.5 Pro | 68 | $20 | 3.40 |
| GPT-5 | 70 | $30 | 2.33 |
| Claude Sonnet 5 | 65 | $15 | 4.33 |
| Claude Opus 5.5 | 73 | $75 | 0.97 |
Opus 5.5 ranks dead last on price-performance. Not surprising — the Max tier was never designed to compete on cost efficiency.
Model Selection Recommendations
Quick guidance by use case:
- Production Agents / long-context Agents: Opus 5.5 is still the top pick — think 120K-line codebases
- General dev work, coding, debugging: Sonnet 5 is plenty; the intelligence gap is easy to close in practice
- Cost-sensitive / high-volume calls: GPT-5 or Gemini 2.5 Pro
- Enterprise contracts, customer-facing SLAs: Opus 5.5 — this is where it actually shines
Editorial Take
Why didn't Opus 5.5 land with the same impact as earlier releases? Three reasons.
First, the previous generation raised the ceiling too high. Opus 4.5 already set the industry bar, so the marginal returns on improvement are visibly shrinking. The HN complaint about "5% gain for 25% more cost" captures the reality.
Second, Anthropic's main battlefield isn't the flagship anymore. In 2026, Anthropic is clearly investing in two directions: Claude Code (CLI tooling) and enterprise API contracts. The Opus Max tier is essentially an anchoring product — a high-priced, high-performance flagship that sets the ceiling for enterprise deals, while actual volume moves through Sonnet and Haiku. Same playbook as luxury car brands with halo models.
Third, competitors are no longer playing catch-up. GPT-5 and Gemini 2.5 Pro have both closed into the 70±3 range, where intelligence gaps become small enough that prompt engineering flattens them. Competition has shifted to ecosystem, Tool Use, and Agent frameworks — none of which a single intelligence score captures.
My prediction: Opus 6.0 will push prices up another 20-30%. But Sonnet and Haiku will hold steady or drop slightly — those are the lines that drive real volume.
Caveat: I haven't seen data on Opus 5.5's vision understanding or multimodal Agent performance. If there's a surprise there, it could upend the "flagship is just an anchor" thesis.
Sources: Anthropic's official release page, Claude Opus 5.5 Intelligence, Performance and Price Analysis on artificialanalysis.ai, Hacker News item #49804316