Do GPT-5.5 vs Claude Benchmarks Matter?

GPT-5.5 vs Claude looks decisive on paper: OpenAI's newest model beat Claude 4.7 on Terminal-Bench 2.0, 82.7% to 69.4% as reported in Lucas Montano's April 2026 video. He argues the comparison matters less than it seems, because the harness around each model shapes daily results more than raw scores.

GPT-5.5 vs Claude: What the Launch Numbers Actually Show

GPT-5.5 vs Claude on published benchmarks favors GPT-5.5, but only inside the specific tests OpenAI chose to show. In Lucas Montano's April 2026 video, Terminal-Bench 2.0 gave GPT-5.5 82.7%, up from GPT-5.4's 75.1%, with Claude 69.4%. Those figures come from the video, not from an independent run, so treat them as vendor-framed numbers.

The comparison table matters for what it leaves out. OpenAI's own internal software-engineering benchmark reportedly scored GPT-5.5 at 73.1 with no published Claude figure beside it, which makes a head-to-head impossible on that row. On OSWorld Verified, a benchmark for computer-use tasks, Montano notes the gains over Opus 4.6 were small, suggesting the models stay roughly balanced outside coding tasks.

A quick summary of what the video reports:

• Terminal-Bench 2.0: GPT-5.5 82.7%, GPT-5.4 75.1%, Claude 69.4%
• Internal OpenAI engineering benchmark: GPT-5.5 73.1%, no Anthropic score published
• OSWorld Verified: modest improvement over Opus 4.6
• Price: about 2x GPT-5.4 per token

None of these numbers is independently verified here, and OpenAI's model announcements are the vendor's own framing. The pattern to notice is selection: every visible benchmark is one the winning vendor controls or highlights.

The Pricing Trade-Off Behind the New Models

The GPT-5.5 vs Claude pricing strategies point in opposite directions, according to the video. OpenAI raised the per-token price: GPT-5.5 costs roughly twice as much as GPT-5.4, justified by more efficient token use. Anthropic took the reverse path with Opus 4.7, letting the model think longer and consume more tokens while keeping the per-token price lower.

These are two bets on where cost should land. OpenAI charges more per unit and spends fewer units. Anthropic charges less per unit and spends more of them. Your actual bill depends on your workload: long reasoning chains favor the cheap-token model, short precise tasks favor the efficient one. Check current rates on the OpenAI pricing page and Anthropic pricing before deciding, since both change without notice.

Why the AI Harness Beats the Model Score

Benchmarks miss the layer that shapes your results: the AI harness. Montano describes it as the operating system wrapped around a model, handling caching, retry logic, tool selection, and when to search. The same model inside Cursor can behave differently from the same model inside OpenAI's Codex because each harness routes, caches, and delegates differently.

The practical consequence is that you no longer fully control what runs. In the video, Montano selected Opus 4.7 in the Claude desktop app, watched it spawn a sub-agent running an older model for a bash step, and found no way to configure that. Benchmarks test the model in isolation; your work tests model plus harness plus sub-agent choices you may not see.

That reframes how to evaluate tools in 2026:

1. Run your real tasks, not leaderboard tasks, in each tool.
2. Compare the same model across harnesses before blaming the model.
3. Verify which model each step actually used when outputs look off.

The Claude Code Effort Bug That Proved the Point

Anthropic confirmed in 2026 that a bug in Claude Code, present since March 4, silently ignored the 'effort high' setting. Users saw effort configured as high while the model actually reasoned at medium depth. Montano reads this as evidence for his earlier prediction that vendors would either limit model capability or add latency to control costs.

The bug sat in the harness, not in the model itself. That is the exact gap benchmarks cannot measure: a model can score at maximum reasoning depth while a production tool quietly runs it lower. For anyone comparing GPT-5.5 vs Claude on paper, the lesson is that configuration behavior in the shipped product can differ from the tested configuration, and only your own testing catches it.

Choosing the Right Model and Tool for the Job

The realistic workflow is becoming model-and-tool triage. Montano suggests using GPT-5.5's strength in search-heavy tasks to draft research and plans, then executing those plans with Opus 4.7. For coding, the choice of harness matters as much: Codex, Claude Code, and Cursor each deliver different results and different costs on the same task.

In the same video, Montano notes Google's statement in 2026 that about 75% of its new code is AI-written, and says his own share is higher. He argues the percentage itself signals little about the job market; what matters is that developers increasingly act as reviewers and directors of AI output. Written AI code that you specified precisely is still your code, in the sense that you own its correctness.

What Tesla FSD in Amsterdam Adds to the Argument

The video's driving segment illustrates the same thesis from another angle. Tesla's Full Self-Driving (Supervised) became available in the Netherlands, and Montano demonstrates it handling Amsterdam's dense bike and pedestrian traffic on cameras alone, with no ultrasonic sensors. The system predicts flow, starting when the car ahead starts, and scores the driver's attention through a cabin camera that can lock out FSD access after repeated distraction.

His point: performance comes from the trained system and its data, not from a spec sheet. Millions of fleet vehicles feeding video back for training is a moat no benchmark captures, and door-to-door parking automation works in the United States but not yet in the Netherlands. Capability is regional, configured, and harness-dependent, just like language models.

FAQ

• How does GPT-5.5 compare to Claude on benchmarks? Per the April 2026 video, GPT-5.5 scored 82.7% on Terminal-Bench 2.0 versus 75.1% for GPT-5.4 and 69.4% for Claude. The comparison is vendor-framed, and one internal OpenAI benchmark had no published Claude Code at all.

• Is GPT-5.5 more expensive than GPT-5.4? The video reports roughly double the per-token price of GPT-5.4, offset by claimed token efficiency. Confirm current rates on OpenAI's pricing page before budgeting.

• Why don't benchmarks predict daily results? Because your output depends on the AI harness, which handles caching, retries, tool calls, and sub-agent model selection. A top-scoring model can behave differently, or run a different model under the hood, inside your editor.

• What was the Claude Code effort bug? Anthropic admitted in 2026 that since March 4, Claude Code displayed 'effort high' while silently reasoning at medium. The flaw lived in the harness layer, not the model.

• Should I stop reading leaderboards entirely? No, but treat them as a filter, not a verdict. Run your own tasks across at least two harnesses, watch which models actually execute each step, and compare cost per completed task.

Turn Video Commentary Into a Written Reference

The core insight of this piece is that claimed performance and delivered performance diverge once real context enters. The same gap exists between watching a creator's take in a twenty-minute video and having it written down where you can quote, link, and search it. Viewers of communities like Gustavo Dev Doido's Crazystack Typescript and Bootcamp do Dev Doido audiences often want that written layer alongside the video.

If you publish commentary, lessons, or explanations on YouTube, Skala Blog turns a video URL into a structured, source-linked article. Paste the link, get the transcription, review the draft, and your argument survives past the algorithm's next benchmark cycle.

Source video