GPT-5.6 vs Claude Opus 4.8: A Split Decision, and a Measured Tie (July 2026)

Most comparison pages treat GPT-5.6 as one model and crown a single winner. Neither is true. We put GPT-5.6 Sol and Claude Opus 4.8 through our own executed nine-task Python benchmark: both scored 9/9, at a measured $4.05 per 1,000 tasks for Opus 4.8 and $4.98 for Sol. On that workload neither one wins. GPT-5.6 also ships as three tiers — Sol, Terra, Luna — and since 2026-07-24 there is a third Anthropic answer nobody was comparing: Claude Opus 5, at the same list price as Opus 4.8 and a measured $5.64. Here is the split decision, with the output-token math the spec tables never bother to compute.

The 30-second verdict

Both models are real, verifiable July-2026 releases. Claude Opus 4.8 is Anthropic hybrid reasoning model, described as built for serious coding and AI agents. GPT-5.6 is not a single model but a three-tier family — Sol, Terra, and Luna — that OpenAI moved to general availability on July 9 2026 across ChatGPT, Codex, the API, and GitHub Copilot after a limited preview that opened June 26.

The honest answer is a split decision, not a coronation:

• On the one benchmark we ran ourselves, nobody wins. GPT-5.6 Sol and Claude Opus 4.8 both scored 9/9 on our nine executed Python tasks. Opus 4.8 was 19% cheaper per task and 0.5 s faster; Sol spent 58 reasoning tokens per call where Opus spent none. Full numbers below.
• Opus 4.8 wins hard software engineering, on third-party evals. It scores 69.2% on SWE-bench Pro versus 64.6% for the strongest GPT-5.6 tier, and it is the strongest computer-use and browser agent at around 84% on Online-Mind2Web. Those are vendor and leaderboard figures, not ours.
• GPT-5.6 Sol wins frontier agentic reasoning, again on third-party numbers: Sol leads OpenAI cited Agents Last Exam eval at 53.6.
• Terra and Luna win on price. They deliver near-frontier coding scores at a fraction of Opus or Sol cost.
• And there is now a third Anthropic option this page originally missed. Claude Opus 5 launched 2026-07-24 at the same $5 / $25 list price as Opus 4.8. We measured it at 9/9 and $5.64 per 1,000 tasks — same score, 39% more money.

If you have only ever thought of this as "GPT vs Claude," start with our broader Claude vs ChatGPT overview, then come back for the model-level detail below.

We ran both: 9/9 each, $4.05 against $4.98

When this comparison first went up we had third-party benchmark numbers for both models and first-party numbers for neither. We now have both sides on our own harness, and the result is the part no competing page can give you: a tie.

GPT-5.6 Sol ran on 2026-07-29 and is priced at 2026-07-29 rates. Claude Opus 4.8 ran in our original 13-model sweep and is priced at 2026-07-17 rates. Those dates differ, which would matter if either list price had moved in between. Neither did: Sol was $5 / $30 per 1M tokens and Opus 4.8 was $5 / $25 on both dates. Rows are ordered by measured cost, cheapest first.

Read the score column first, then stop expecting it to decide anything. Opus 4.8 and Sol both went 9/9 with no misses. The separation is entirely on the other columns:

• Cost. $4.05 against $4.98 per 1,000 tasks — Opus 4.8 is 19% cheaper, or put the other way, Sol costs 23% more for the identical result. Two things push in the same direction: Sol output lists 20% higher ($30 against $25 per 1M) and Sol spent reasoning tokens that Opus 4.8 did not.
• Latency. 6.1 s against 6.6 s. A 0.5 s edge to Opus 4.8. That is marginal, and on a nine-task sample we would not build a plan on it.
• Reasoning tokens. 58 per call for Sol, 0 for Opus 4.8. Reasoning tokens bill at the output rate, so a model at zero has no thinking-token surcharge and no cost variance from prompt to prompt. 58 is a small trace by the standards of this harness — Gemini 3.1 Pro averaged 1,094 on the same nine tasks — but it is not zero.

The honest framing: neither model wins this workload decisively. Opus 4.8 takes a narrow cost and latency edge; Sol matches it on every task. If your work looks like short, clearly specified Python functions, that difference is a rounding error next to what routing or caching would save you, and a 19% gap is not a reason to migrate a stack.

It is also worth saying why a tie is unsurprising. This harness has now run 23 models — the original sweep was 13 and the rest were run later under identical settings — and 20 of the 23 scored 9/9. Nine bounded functions do not separate frontier models. They separate their bills, which range from $0.10 to $14.70 per 1,000 tasks across those 23 runs, a 147x spread on a mostly flat score column. That is the finding this page exists to hand you, and it is the same finding as the wider AI coding ranking.

The third option: Claude Opus 5

This comparison was written when Opus 4.8 was the newest Anthropic flagship. It no longer is. Claude Opus 5 launched 2026-07-24, and a reader choosing today has a third answer.

What Anthropic and the launch coverage report — third-party, not measured by us: a 1M-token context window with no beta header and no long-context price premium, up to 128K output tokens, adaptive thinking on by default, and a five-level effort setting including a new xhigh mode. Anthropic positions it as approaching Claude Fable 5 intelligence at half the price, and prices it identically to Opus 4.8. See anthropic.com/claude/opus and the launch coverage for the specifications. We have never run Claude Fable 5 through this harness, so we take no position on the Fable comparison.

What we did measure, on 2026-07-30: 9/9, $5.64 per 1,000 tasks, 5.3 s average, 6 reasoning tokens. And here is the counterintuitive part.

Opus 5 does buy one thing on this harness: it is the fastest of the three at 5.3 s, against 6.1 s for Opus 4.8 and 6.6 s for Sol. So the choice among the three, on bounded Python work, reads: Opus 4.8 cheapest, Opus 5 fastest, Sol neither, all three perfect. A newer generation at the same sticker price costing more per completed task is exactly the trap a per-token price list cannot show you, and the reason we rank by cost per completed task instead. If latency is what you are buying, note that Anthropic also ships a Claude Opus 5 Fast variant, which we measured at 9/9 in 3.4 s at $10.20 per 1,000 tasks priced 2026-07-30 — 1.8x the money for 1.9 s, and 2.5x what Opus 4.8 cost.

What none of this tells you is whether Opus 5 is better at the work Anthropic actually built it for. Adaptive thinking, an xhigh effort level and a 1M-token window are features for long, ambiguous, multi-step jobs. Our nine tasks are none of those. A tie here means the test did not reach the difference, not that there is none — see what we did not measure.

Pricing: output is the number that matters

Agentic coding loops are output-heavy — the model writes far more tokens than it reads once tools, plans, and diffs start flowing. That makes the output price, not the headline input price, the number that actually drives your bill. Here is the verified July-2026 picture, side by side.

The quiet surprise: Opus 4.8 output ($25/1M) is cheaper than GPT-5.6 Sol output ($30/1M) — and third-party leaderboards put Opus above Sol on SWE-bench Pro. So on the hardest coding work, the higher-scoring model is also the cheaper one per output token. Our own run agrees on the cost half of that claim and finds no gap at all on the score half. Context and max-output are a wash: all four offer a 1M-token window and 128,000 max output tokens. GPT-5.6 also carries a February 16 2026 knowledge cutoff across every tier.

One caution before you copy any of these figures elsewhere: several aggregator pages carry hallucinated GPT-5.6 specs — one listed tiers at $4/$16, $8/$32, and $14/$56 with mismatched 512K and 1.5M context sizes. Those contradict the primary numbers from OpenAI, Simon Willison, and MarkTechPost. Trust only the table above, and see the dated caveat on the Terra and Luna rates in the tier section.

Benchmarks: coding vs agents

On the coding number that is most consistently sourced — SWE-bench Pro, which measures real repository diffs — Opus 4.8 leads the active leaderboard. It climbed to 69.2% from 64.3% on Opus 4.7; if you want that generational jump in detail, see Claude Opus 4.8 vs 4.7. All three GPT-5.6 tiers cluster just below it.

Flip to open-ended agentic reasoning and the ranking flips too. GPT-5.6 Sol leads OpenAI cited Agents Last Exam at 53.6, and the family introduces Programmatic Tool Calling — the model writes JavaScript that runs in an isolated V8 runtime with no network access, exposed in the Responses API. On terminal automation, GPT-5.6 is strong across the board (Terminal-Bench 2.1: Sol 88.8, Terra 87.4, Luna 84.7). Treat every one of these as vendor-reported and eval-specific — they are snapshots, not laws, and these leaderboards move within weeks.

Opus 4.8 SWE-bench Verified is reported inconsistently across leaderboards (figures range from the low-80s to about 88.6%), which is exactly why we lead with the more stable SWE-bench Pro number instead.

The gap between these third-party rankings and our own run is the useful thing here, not a contradiction. SWE-bench Pro puts 4.6 points between Opus 4.8 and Sol on multi-file repository diffs. Our harness puts zero points between them on nine self-contained functions. Both are true, and which one predicts your bill depends on which of those two jobs you are actually paying for.

Sol vs Terra vs Luna

Collapsing GPT-5.6 into one row is the mistake most comparisons make. The tiers are built for different jobs:

• Sol — flagship frontier reasoning for complex agentic and scientific work. Coding Agent Index v1.1: 80. This is the tier we measured: 9/9, $4.98 per 1,000 tasks, 6.6 s, priced 2026-07-29. Detail in our GPT-5.6 Sol review.
• Terra — balanced mid-tier for production, at $2.50/$15 on launch-day pricing. Coding Agent Index: 77.4. Not run on our harness.
• Luna — lightweight, high-throughput for classification, routing, and volume, at $1/$6 on launch-day pricing. Coding Agent Index: 74.6. Not run on our harness.

OpenAI pitches Terra and Luna as outperforming a rival at roughly one-sixteenth the cost. On cache economics, aggregator sources report cache writes billing at 1.25x uncached input with cached reads getting a 90% discount — useful if true, but treat it as unconfirmed until you see it on OpenAI own pricing page.

One dated caveat on the tier prices. The $2.50/$15 and $1/$6 figures above are the launch-day numbers from the primary sources. Our own catalogue snapshot on 2026-07-29 lists Terra at $1.25 / $7.50 and Luna at $0.50 / $3 — below the launch rates. Prices move: 49 of roughly 396 models in our catalogue changed price in the twelve days to 2026-07-29, which is why every figure on this page carries a date. Check live pricing before you budget, and see how fast list prices actually move. If you are weighing whether to move up from the last generation, our GPT-5.6 vs GPT-5.5 and GPT-5.6 launch breakdown cover the delta; the older GPT-5.5 model page has its specs.

A computed cost example

Spec tables never do the arithmetic, so here it is. Assume a realistic agentic coding session — 8K input and 2K output tokens per turn — run for 1,000 turns (8M input, 2M output total), at the launch list prices in the pricing table above:

Two facts jump out. First, for the hardest work, Opus 4.8 is both the higher-scoring option on third-party evals and the cheaper of the two frontier models — $90 versus $100 for Sol. Second, Luna does the same 1,000 turns for $20, roughly a fifth of Opus, at a SWE-bench Pro score only ~6.5 points lower. That gap is the whole argument for routing: pay Opus prices only where the extra points earn their keep. In our 2026-07-29 catalogue snapshot Terra and Luna are cheaper still, which only widens it.

Now the part that table hides. It assumes both models emit the same number of tokens. Ours did not. On the nine tasks we actually executed, Opus 4.8 came in 19% cheaper per task than Sol — $4.05 against $4.98 — not the 10% an equal-token assumption predicts. Sol pays a 20% higher output rate and spent 58 reasoning tokens per call where Opus 4.8 spent none, and both effects land on the same side. Assumed-token cost tables systematically understate the gap between a reasoning model and a direct-answer one. To run the arithmetic on your own token mix rather than ours, use the cost calculator.

Which one for which task

The frame that beats a single-winner verdict:

• Bounded, clearly specified code generation → Claude Opus 4.8, or honestly something much cheaper. On our nine tasks it tied Sol and Opus 5 at 9/9 for the least money of the three, and 20 of the 23 models we have run also scored 9/9 — including one at $0.10 per 1,000 tasks. See the cheap coding model roundup before you pay frontier rates for function-level work.
• Hardest agentic coding, browser/computer use, real repo diffs → Claude Opus 4.8. Highest third-party SWE-bench Pro of these four, top computer-use score, and cheaper output than Sol.
• Frontier open-ended agentic reasoning, science-flavored tool loops → GPT-5.6 Sol, especially if Programmatic Tool Calling fits your stack.
• Production workloads at scale → GPT-5.6 Terra — near-frontier coding at half the cost, or less at current catalogue rates.
• High-volume classification, routing, batch → GPT-5.6 Luna — $20 per 1,000 turns in the example above.
• Latency-sensitive work, and long ambiguous jobs → consider Claude Opus 5 at the same $5 / $25 as Opus 4.8. It was the fastest of the three we measured at 5.3 s, and its 1M-token context, adaptive thinking and xhigh effort level are aimed squarely at the work our harness cannot see. Just go in knowing it cost 39% more than Opus 4.8 for the same 9/9 on bounded tasks.

One hedge worth stating plainly, now with a name on it: Opus 4.8 is no longer the top of Anthropic lineup. Claude Opus 5 shipped 2026-07-24, and Claude Fable 5 sits above both — we have never run Fable 5, so we make no claim about it either way. Do not read "Opus 4.8 leads SWE-bench Pro" as "the top of Anthropic entire lineup." Within the models this comparison is about, the routing table above holds. And because DataLLM Lab keeps all of them on one key, you do not have to pick once — you pick per request.

How our numbers were produced

Every first-party figure on this page comes from one harness: 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 list price on a stated date — 2026-07-29 for GPT-5.6 Sol, 2026-07-17 for Claude Opus 4.8, 2026-07-30 for Claude Opus 5. Both Sol at $5 / $30 and Opus 4.8 at $5 / $25 were unchanged across those dates, so the mixed pricing dates do not distort the comparison. List prices do move, though — 49 of roughly 396 catalogue models changed price in the twelve days to 2026-07-29 — so an undated cost figure is not a fact.

Calls go through OpenRouter 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; GPT-5.6 Sol, Claude Opus 5 and eight others were run later on the same harness under the same settings, bringing the total to 23. Where this page says 23 models or 20 of 23, that is the combined set. Full method on the methodology page; the whole run is in the coding cost benchmark.

What we did not measure

Not measured at all: long-context reasoning, multi-file refactoring, agentic or multi-turn tool use, any non-Python language, and vision. The harness is single-turn. It does not run agents and it does not use tools. Almost everything these two flagships are sold on sits in that list — which is exactly why we do not present a 9/9 tie as "these models are equivalent."

A known scoring 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 three models on this page was affected — all scored 9/9 — but the ceiling is part of the harness.

Nine tasks is nine data points. A 9/9 is a clean result on this set, not a claim about a distribution, and a 0.5 s latency difference on nine calls is not a performance model. We did not run anything twice, we did not vary the prompts, and we did not buy any model a retry.

Not tested, and never claimed as ours: GPT-5.6 Terra, GPT-5.6 Luna, Claude Fable 5, GPT-5 nano, 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. Every Terra and Luna number on this page is third-party. If a first-party figure for any untested model appears anywhere on this site, it is an error.

FAQ