Claude vs GPT-5 in 2026: We Ran Nine Tiers Across Both Families (All 9/9, $0.94 to $8.83)

Claude and GPT-5 are the two leading frontier families of 2026, and most comparisons end in a verdict nobody can check. So we ran them. Five Claude tiers and four GPT-5 tiers went through the same executed nine-task Python benchmark in July 2026, and all nine scored 9/9. On work that size the family choice is not a capability choice at all — it is a cost and latency choice, and there the spread is $0.94 to $8.83 per 1,000 tasks. That result is narrow on purpose: nine short, self-contained functions say nothing about long context, multi-file refactoring or agentic tool use, which is exactly where the two families genuinely diverge. This guide gives you our measured numbers, the independent third-party benchmarks, modeled monthly costs, and worked scenarios for picking by job.

The short answer

On short, well-specified coding work, both families are the same model. We ran nine tiers across Claude and GPT-5 on the same executed benchmark and every one returned 9/9. What separates them there is money: Claude Haiku 4.5 at $0.94 and GPT-5.4 at $1.69 per 1,000 tasks, against Claude Opus 5 at $5.64 and GPT-5.5 at $8.83 for the identical result.

Where the families genuinely diverge is everything our harness does not reach — long context, multi-file refactoring, agentic tool use, non-Python work, writing quality. There the third-party picture still splits: independent SWE-bench Verified puts Claude Opus ahead on repository bug-fixing, Terminal-Bench puts GPT-5 ahead on shell-driving agents. For most teams the real answer stays both, routed by task — with the cheap tier doing far more of the work than its price suggests.

What we measured: nine tiers, one score

Across July 2026, five Claude tiers and four GPT-5 tiers went through the same executed nine-task Python harness. The model gets a function signature and a prose spec, never sees the assertions, and the returned code is executed against them. One scored attempt each, temperature 0.

All nine scored 9/9. No misses, in either family, at any tier. Rows are ordered by measured cost per 1,000 tasks, cheapest first. The last column is the date of the list price each cost was computed at; Sonnet 5, Opus 4.8 and GPT-5.5 came from our original 13-model sweep, the other six ran later on the same harness on the dates shown.

The score column never moves and the cost column moves 9.4x. That is the finding. The two flagships people actually reach for cost 6.0x and 9.4x what Haiku 4.5 cost for the same nine results — Claude Opus 5 at $5.64 and GPT-5.5 at $8.83 against $0.94.

Three things in that table are invisible on any rate card, because they come from token counts rather than prices:

• GPT-5 mini lists 10x cheaper than GPT-5.4 on input, 7.5x on output — and measured within 10% of it. $0.25 / $2 per 1M against $2.50 / $15, yet $1.53 against $1.69 per 1,000 tasks. GPT-5 mini spent 555 reasoning tokens per task; GPT-5.4 spent zero. Reasoning tokens bill at the output rate, and here they ate almost the entire list-price advantage. GPT-5.4 also returned in 3.6 s against 15.2 s. The two tiers are covered head to head in Claude Haiku vs GPT-5 mini.
• GPT-5.6 Sol and GPT-5.5 share a $5 / $30 list price and measured $4.98 against $8.83 — 1.8x apart on token count alone, same 9/9. See the GPT-5.6 Sol run and the GPT-5.5 run.
• Claude Opus 5 and Claude Opus 4.8 share a $5 / $25 list price and measured $5.64 against $4.05 — 39% apart, again from tokens returned, not from a repricing. Detail in the Opus 5 review.

Cross-family, at both ends of the ladder, Claude was the cheaper option on this workload: $0.94 against $1.53 at the cheap tier, $5.64 against $8.83 at the flagship. On latency GPT-5.4 was the quickest of the nine at 3.6 s, one tenth of a second ahead of Claude Haiku 4.5, and GPT-5 mini was the slowest at 15.2 s. Latency and cost do not rank the same way, which is the same disagreement we found across the whole field in the AI coding ranking.

These nine are part of a larger set. Our original sweep was 13 models run in one sitting, of which 10 scored 9/9; ten more have since run on the same harness under the same settings, and 20 of the 23 scored 9/9. Across all 23 the measured spread is $0.10 to $14.70. Full run: the LLM coding cost benchmark.

What nine short functions cannot tell you

Read the section above with its scope attached, because the scope is small. The harness runs nine short, self-contained Python functions. Every task fits in a few hundred tokens, is single-turn, and has an unambiguous spec.

It does not measure: long-context work, multi-file refactoring, agentic or multi-turn tool use, any non-Python language, vision, translation, or writing quality. It does not run agents and it does not use tools.

That list is not a footnote — it is the entire commercial argument for the expensive tiers, and it is where the two families actually differ. Anthropic sells Opus for sustained multi-step reasoning and long context; OpenAI sells GPT-5.5 and the Codex variants for agentic execution. Our nine tasks sit far below the level at which either claim starts to be tested. A tie here means the test did not reach them, not that the tiers are interchangeable.

So do not read this page as "just use Haiku". Read it as: for bounded, clearly specified code generation, the cheap tier of either family did everything the flagship did, and the burden of proof sits on the tier that wants 6x to 9.4x. Find where your own workload stops being easy — that is the only place the tier question has an answer, and it is not a place a nine-task benchmark can find for you.

Side by side

Note the row that flips: GPT-5 mini has the cheaper list price — 4x on input and 2.5x on output — and Claude Haiku 4.5 has the cheaper measured cost by 1.6x. Reasoning tokens are the whole difference. Per-token rate cards cannot show you that, which is the reason we run the tasks.

Third-party benchmarks

Our harness answers a narrow question. For the broader one, the independent boards split. On the independent SWE-bench Verified board, Claude Opus 4.8 (88.6%) leads GPT-5.5 (82.6%) on real-world repository bug-fixing:

The nuance behind the headline: SWE-bench measures bug-fixing against real repositories, where Claude leads; Terminal-Bench measures command-line agent execution — running commands, navigating a shell, recovering from errors — where GPT-5 leads (~82.7%). So "better at coding" genuinely depends on whether your work is patching code or driving a terminal. Treat a single benchmark as one data point, not the verdict, and weight the one that matches your workload.

Read against our run, the two sources are consistent rather than contradictory. SWE-bench tasks are large, multi-file and ambiguous; ours are small, single-file and precise. The families separate on the first kind of work and tie on the second. That is the practical shape of the whole comparison: the harder the task, the more the family choice matters — and the more of your traffic is easy, the less you should be paying for it. Rankings across the field are compared in best coding LLM 2026.

What they cost to run

List prices are close enough at the flagship tier that they rarely decide it — but the tier you pick within each family changes the bill by an order of magnitude. Here is the modeled monthly cost across five workloads, from list price and stated token volumes:

Opus 5 and GPT-5.5 land within 4% to 18% of each other across these five workloads — close enough that list price rarely decides the flagship; GPT-5.4 is the cheapest frontier option on list price. But the real lever is dropping to a cheap tier where quality allows — GPT-5 mini models a chatbot at $34 against Opus at $500.

Two cautions on that table. First, it assumes a fixed token volume, and our measured run shows models with identical list prices consuming very different token counts for the same work — Opus 5 versus Opus 4.8 differed by 39%, GPT-5.5 versus GPT-5.6 Sol by 1.8x. Second, reasoning tokens are output tokens: GPT-5 mini's 555 per task on our harness is why its modeled $34 will understate a reasoning-heavy workload. Run your own mix through the cost calculator rather than trusting either table as a quote. Prices also move — 49 of roughly 396 catalogue models changed price in the twelve days to 2026-07-29.

Where Claude wins

• Cost per completed task on bounded coding work — measured: cheaper than the GPT-5 tier at both ends, $0.94 vs $1.53 and $5.64 vs $8.83 per 1,000 tasks.
• Predictable billing — measured: four of the five Claude tiers we ran emitted zero reasoning tokens and Opus 5 emitted six, so cost per call barely varies with prompt awkwardness. Every GPT-5 tier except 5.4 emitted a reasoning trace.
• Coding quality on large, ambiguous tasks — third-party: leads independent SWE-bench Verified (vals.ai, June 2026). Not measured by us.
• Planning & instruction-following — widely preferred for staying on-task across complex, multi-step instructions. Not measured by us.

Where GPT-5 wins

• Latency, at the mid tier — measured: GPT-5.4 was the fastest of the nine at 3.6 s, and it emitted zero reasoning tokens doing it. Note this does not extend to GPT-5 mini, which was the slowest of the nine at 15.2 s. Tier comparison in GPT-5.4 vs GPT-5.5.
• Terminal agents — third-party: leads Terminal-Bench; the Codex variants are tuned for coding agents that drive a shell. Not measured by us.
• Ecosystem — the native OpenAI format with the widest tooling, libraries, and integrations.
• Tier range — from nano to Pro, a wider price ladder for fine-grained cost control.

Worked scenarios

Which to pick by job

How our numbers were produced

Nine executed Python tasks: two_sum, valid_parentheses, merge_intervals, roman_to_int, lcs_len, flatten, top_k_words, token_bucket, parse_csv_line.

The model receives a function signature and a prose spec. It never sees the assertions. The returned code runs against hidden assertions in an isolated python3 -I subprocess with a 12-second timeout. Every assertion passes or the task fails — no partial credit, no human grader, no LLM judge. Temperature 0, max_tokens 4000, one scored attempt per task; the harness retries only on an API error, never on a wrong answer.

Cost is derived, not invoiced. It is the token counts the API reported multiplied by that model's list price on the date shown in the table — 2026-07-17 for Sonnet 5, Opus 4.8 and GPT-5.5; 2026-07-29 for Haiku 4.5, Sonnet 4.6, GPT-5 mini, GPT-5.4 and GPT-5.6 Sol; 2026-07-30 for Opus 5. Write it down as measured cost, not a bill. Prices move, so recompute against current rates before acting on any figure here.

Calls go through OpenRouter's OpenAI-compatible endpoint, deliberately not through the DataLLM Lab gateway. Nothing on this page depends on our infrastructure and you do not have to be our customer to reproduce it.

Our original sweep was 13 models run in one sitting, 10 of which scored 9/9. Ten more have since run on the same harness under the same settings, bringing the total to 23, of which 20 scored 9/9. The core sweep was and remains 13.

Not tested, and never claimed as ours: Claude Fable 5, Claude Opus 4.7, GPT-5 nano, the Codex variants, Ollama or any locally-run model, Grok 3, Grok 4, Grok 4.20, grok-build-0.1, VibeThinker, and any Gemini other than 3.6 Flash and 3.1 Pro. One known artefact: the 4,000-token ceiling can truncate a very verbose model mid-answer and a truncated answer scores as a miss. None of the nine models on this page were affected.

FAQ