AI Model Reviews

Claude Opus 5.5 vs GPT-6 Sol: Two Models, Two Different Jobs

Claude Opus 5.5 and GPT-6 Sol may occupy nearby places in their companies' product lines, but Peter's tests suggest they should not occupy the same place in a working AI stack. Sol behaves like an efficient production model: quick enough to iterate with and cheap enough to run repeatedly. Opus behaves more like a premium production team: slow and expensive in these stress tests, but capable of building artifacts that the cheaper run did not approach.

The short answer: use GPT-6 Sol to explore, implement, test, and iterate. Escalate to Claude Opus 5.5 when the final artifact is valuable enough to justify hours of agent time and tens of dollars in model usage. Peter's strongest Opus results are remarkable, but they do not establish it as the best model for routine coding, nor do they make every expensive run worthwhile.

Watch Peter's Full Opus 5.5 and GPT-6 Sol Test

Source and credit: Claude Opus 5.5 | First impressions and comparison with GPT-6 Sol, published by Arena AI on 23 September 2026. Peter's reported times, token totals, costs, line counts, and visual judgments come from this creator test. Anthropic and OpenAI are the sources of record for current model pricing and availability.

Why the Head-to-Head Comparison Breaks

Both models were run at maximum reasoning against the same creative prompts. That sounds controlled until the models choose radically different amounts of work. In the Golden Gate Bridge task, Sol returned a compact, competent 3D interpretation. Opus built a far richer environment, including water, landscape, city detail, and additional structures. One answer was a prototype; the other behaved like a finished showcase piece.

That difference creates a category problem. A model that spends ten minutes and around $1.50 is competing for frequent use. A model that spends several hours and $50-$90 is competing for a smaller number of high-value deliverables. Comparing only beauty makes Sol look weak. Comparing only price makes Opus look irrational. The useful metric is cost per accepted outcome at the quality level the job actually requires.

What Peter Reported in the Creative Builds

BuildGPT-6 SolClaude Opus 5.5What the test suggests
Golden Gate BridgeAbout 5M total tokens, $1.70, 858 lines146.8M total tokens, $61, about 5,500 linesOpus produced dramatically more detail at dramatically higher cost
StonehengeLower-cost comparison shown in the videoAbout 7 hours, $61, 140M total tokens, roughly 3,500 linesOpus treated a single prompt like a long production assignment
CastleAbout 10 minutes, under $1.50, 359 linesAbout $90, 150M total tokens, roughly 5,000 linesThe outputs did not target the same finish level
OctopusCompetent but visibly simplerAbout $50 and much richer movement and atmosphereCreative direction, not just geometry, separated the results
London through timeNot the featured comparisonAbout $65 and 7,000 linesOpus sustained one historical visual system across multiple eras
Trojan Horse vasePeter cites a roughly $0.76 alternativeAbout $35 and 1,500 linesThe expensive result became a coherent animated story, not merely a scene

These figures are useful observations, not a benchmark dataset. Peter ran a collection of ambitious prompts in one agent environment and judged the visual results himself. He repeatedly noted uncertainty about why some runs lasted four to seven hours and whether the machine or workflow contributed. There were no repeated trials, blind reviewers, fixed tool-call budgets, or normalized stopping conditions.

Read the Enormous Token Totals Carefully

A figure such as 146.8 million tokens is easy to misread. It does not mean Opus wrote 146.8 million visible tokens of source code. The final artifact contained thousands of lines, not millions of pages. In an agent loop, a displayed total may accumulate repeated context, cached reads, reasoning, tool results, test output, file inspection, and many turns of work.

The cost also shows that the displayed total cannot be treated as entirely new output at Anthropic's $20-per-million output rate. The right conclusion is narrower: the Opus workflow performed a very large amount of agent activity and incurred a reported bill of $61. To understand why, the experiment would need separate counters for fresh input, cache reads and writes, output, tool calls, retries, and wall-clock waits.

For your own tests: record both the provider invoice and the agent interface totals. They answer different questions. The invoice tells you what was billed; the interface can reveal how much context and repeated work moved through the system.

Official Pricing Makes the Roles Clearer

Standard API rateInput / 1MCached input / 1MOutput / 1M
Claude Opus 5.5$4$0.20 cache reads$20
GPT-6 Sol$2$0.20$10

At standard short-context rates, Sol is half the list price of Opus for fresh input and output. Anthropic says Opus 5.5 costs about 40% less than Opus 5 on typical workloads, generates output more than 30% faster, and performs at the level of Fable 5.1 on most work. OpenAI positions Sol as the balance of intelligence and cost beneath Astra, with a 1.05-million-token context window and reasoning levels from none through max.

Those list prices are only the beginning. Long-context thresholds, cache behavior, fast processing, batch or flex processing, regional inference, tool charges, and failed runs all change the task total. For long agent jobs, the stopping rule may matter more than the nominal per-token difference.

Creative Coding Is the More Interesting Story

The strongest Opus examples were not ordinary websites. They were code-generated visual narratives: London changing across two millennia, the Truman Show reaching the edge of its painted world, a brush-stroke interpretation of a familiar dance, Icarus as an animated children's-book scene, the Berlin Wall, and the Trojan Horse painted around a rotating Greek vase.

The Trojan Horse prompt is revealing because it is short. Peter asked for a 20-second HTML animation around an ancient Greek vase, with the horse entering the city, night falling, and hidden warriors emerging, all drawn in code without external assets. The result's value came from interpretation, sequencing, visual language, and implementation working together.

This is where a high-capability coding model can become a new production medium. The model is not only implementing a specified interface; it is proposing art direction and then constructing the rendering system needed to express it. That can justify premium inference when the output becomes a campaign piece, interactive explainer, museum experience, pitch artifact, or creative prototype that would otherwise require several disciplines.

Does Opus 5.5 Make Fable 5.1 Hard to Justify?

Peter's strategic theory is that Opus now competes more directly with Anthropic's own Fable tier than with Sol. Anthropic's launch language supports part of that intuition: the company says Opus 5.5 performs at the level of Fable 5.1 on most work. But "most work" is not every workload, and Peter's visual examples do not isolate the tasks where Fable may retain an advantage.

The responsible conclusion is to test tier boundaries rather than assume them. If Opus matches Fable on your hardest accepted outputs at a lower total cost, route those tasks down. If Fable wins enough high-value cases to repay its premium, keep it as an escalation tier. Product names are not a routing policy.

Opus Still Showed Ordinary Agent Failures

Peter explicitly warned against calling Opus 5.5 the best coding model on the basis of these demos. In separate coding work, it overwrote changes, returned a Python command instead of launching the application itself, and behaved in ways he found unexpectedly rough. Those failures matter because a visually impressive output can distract from repository safety and execution discipline.

Longer runs also create a larger failure surface. Before granting a model several hours, protect the workspace with version control or an isolated worktree, define files it may change, cap spend and attempts, require progress checkpoints, and preserve a runnable state after each stage. Ambition should increase observability, not reduce it.

A Practical Three-Lane Model Router

LaneDefault modelGood fitsPromotion rule
ExploreGPT-6 Sol at medium or high effortConcepts, architecture, scaffolds, small features, variants, debuggingPromote when the accepted result needs depth or polish Sol repeatedly misses
ProduceGPT-6 Sol high/max or Opus 5.5 with a bounded briefImportant cross-file work, interactive prototypes, client-facing assetsChoose by expected reviewer time and value of the output, not model prestige
ShowpieceOpus 5.5 with staged approvalsHigh-value creative coding, complex visual narratives, difficult one-off artifactsStop or reduce scope when checkpoints do not justify the remaining budget

The economical pattern is draft with Sol, identify the part where capability is actually limiting the outcome, and send only that bounded stage to Opus. A $60 run can be sensible for a valuable final artifact. It is wasteful as an unmonitored first attempt at a task whose definition of done is still changing.

How to Evaluate the Models on Your Own Work

  1. Choose two task classes. Use one routine implementation task and one genuinely ambitious deliverable.
  2. Freeze the brief. Keep files, tools, permissions, model effort, and acceptance criteria consistent.
  3. Add staged stopping points. Review plan, first runnable version, and final polish separately.
  4. Record the complete cost. Capture invoice cost, displayed tokens, elapsed time, retries, tool failures, and reviewer minutes.
  5. Score accepted value. Evaluate correctness, visual or functional quality, maintainability, and how much human repair remains.
  6. Route by threshold. Keep Sol as the default unless Opus improves the accepted result enough to repay its extra cost and delay.

Peter's best closing advice is also the most useful: do not test a capability jump only with yesterday's small prompts. Try one more ambitious task. Just give that ambition a budget, checkpoints, and a rollback path.

Video Chapters

TimeChapterTimeChapter
00:00Introduction to Opus 5.500:16Why the comparison breaks down
01:00Golden Gate Bridge comparison02:02Tokens, cost, and code volume
03:29Seven-hour Stonehenge run04:35The $90 and $1.50 castles
05:14Octopus scene07:41Opus versus Fable positioning
08:12London across 2,000 years10:48Tower of Babel and RL theory
14:20Manhattan and Istanbul routing15:53Animated HTML as a strength
16:18Truman Show scene17:34Painted animation
18:23Freddie Mercury at Wembley19:38Icarus visual story
20:22Fall of the Berlin Wall20:46Trojan Horse vase
23:19Final verdict and caveats

Sources, Prompts, and Useful Links

The video was published on 23 September 2026. Official model and pricing pages were checked on 24 September 2026. Model behavior, pricing, availability, context thresholds, and processing tiers can change.

Common questions

Is Claude Opus 5.5 better than GPT-6 Sol?
Not for every job. Peter preferred Opus 5.5 by a wide margin on ambitious visual and animated HTML builds, but those runs could take hours and cost tens of dollars. GPT-6 Sol usually finished the compared examples in minutes for roughly one or two dollars. The useful choice depends on the value of the output and the cost of waiting.
How much do Claude Opus 5.5 and GPT-6 Sol cost?
Anthropic lists standard Opus 5.5 pricing at $4 per million input tokens and $20 per million output tokens, with cache reads at $0.20 per million. OpenAI lists standard GPT-6 Sol pricing at $2 per million input tokens and $10 per million output tokens, with cached input at $0.20 per million. Long-context, fast, regional, batch, and other processing tiers can change the rates.
Did Opus 5.5 really use 146.8 million tokens for one build?
That is the total shown in Peter's Arena workflow for the Golden Gate Bridge run. It should not be interpreted as 146.8 million newly generated output tokens. Agent interfaces can aggregate repeated context, cache traffic, reasoning, and tool-loop activity. Peter also said he was unsure whether the machine or run behavior contributed to the extreme totals and durations.
Does Opus 5.5 replace Claude Fable 5.1?
Anthropic says Opus 5.5 performs at the level of Fable 5.1 on most work, while Peter argues that his visual tests make the tiers feel unusually close. That does not prove Fable is obsolete. Teams should compare accepted-result rate, latency, cost, and review effort on their own hardest work.
What is Claude Opus 5.5 especially good at in these tests?
Peter found its strongest advantage in ambitious code-generated visual experiences: detailed 3D scenes, historical transitions, and animated HTML stories with coherent art direction. That is a narrow but revealing test category, not a universal coding benchmark.
Which model should I use as the default?
Use GPT-6 Sol for routine implementation, exploration, repeated iterations, and cost-sensitive agent work. Escalate selected tasks to Opus 5.5 when a high-value artifact needs substantially more depth, polish, or creative interpretation and the budget can support a longer run.
Share
X LinkedIn Reddit
Build Yours

Want a system
like this one?

Book a free 30-minute call. We map your situation, identify the highest-impact automation, and figure out if we are a fit.

Book Free 30-min Call