GPT-5.6 vs Claude 5: why the benchmarks disagree

Results available on July 30, 2026 do not produce one clear winner between GPT-5.6 and Claude 5. On Datacurve’s shared-agent DeepSWE, Opus 5 and Sol are statistically tied at 74% and 73% [1], while Terra matches Fable 5’s 70% for less than a quarter of the cost per task [1]. My choice changes with the job and budget.

Which model would I use for each coding job?

I would use Terra for high-volume routine work, Sol for a clear goal, Opus 5 for an unfamiliar problem and Fable 5 for careful repository-wide changes. The model name alone is not enough because the task and effort setting can change both quality and cost.

The work	My pick	Effort
Routine endpoints, mappings, small tests, config	Terra	medium or high
A scoped pull request with clear acceptance criteria	Opus 5	medium
Implementing a known plan, fixing a reproducible bug	Sol	high
Root-causing an unfamiliar failure	Opus 5	high or xhigh
Test strategy for a feature that matters	Fable 5	xhigh
Repo-wide refactoring and interface migrations	Fable 5	xhigh
Long terminal and debugging sessions	Sol or Opus 5	high or xhigh
Open-ended engineering with no known solution	Opus 5	xhigh

The measured patterns behind those picks are clearer than any universal model ranking. Sol alone improves its patch quality at every effort step on Cognition’s FrontierCode [15], while Fable performs best at xhigh on that benchmark and Scale Labs’ Test Writing [15] [19]. Opus 5 needs its effort matched to the work, with medium for a scoped change on FrontierCode [15] and xhigh for an open problem in Anthropic’s FrontierBench results [16].

The vendors’ guidance fits those patterns. OpenAI calls medium a balanced starting point [3], while Anthropic defaults to high and recommends max only when the task justifies the extra tokens [6].

How GPT-5.6 and Claude 5 differ on price and effort

Terra is the price exception: OpenAI charges half the Sol rate, while OpenAI and Anthropic give all four models roughly one million tokens of context [2] [4]. Effort settings can change the cost and result even more, but the two vendors’ labels are not directly comparable.

The vendors give these models different jobs. OpenAI positions Sol as its flagship GPT-5.6 model and Terra as the lower-cost option, with the plain gpt-5.6 alias pointing to Sol [2] [3]. Anthropic recommends Opus 5 for complex coding and enterprise work, while Fable 5 sits at the top of its range for long-running agents [4].

Their list prices reinforce that split. Sol costs $5 per million input tokens and $30 per million output tokens, while Terra costs exactly half as much at $2.50 and $15 [2]. Opus 5 is priced at $5 and $25, compared with $10 and $50 for Fable 5 [5]. Anthropic also lists a faster Opus 5 mode in research preview, but charges Fable 5 rates for it [5].

Spec	Sol	Terra	Opus 5	Fable 5
Input, $/MTok	5.00	2.50 (Best value in this row)	5.00	10.00
Output, $/MTok	30	15 (Best value in this row)	25	50
Cache read, $/MTok	0.50	0.25 (Best value in this row)	0.50	1.00
Context window	1.05M	1.05M	1M	1M
Effort levels	6, none to max	6, none to max	5, low to max	5, low to max

Figure 1 makes the practical difference clear: Terra is the only lower-cost model in the group. The context windows do not separate them.

Effort settings matter even more than the list prices, but the two scales work differently. GPT-5.6 offers six settings from none to max, with OpenAI recommending medium as a balanced starting point [3]. Claude offers five, defaults to high, and points demanding coding or agent work toward xhigh; Anthropic reserves max for tasks that justify unrestricted token use [6]. Because these labels are not standardized units, OpenAI’s high may represent a different amount of compute from Anthropic’s.

What more effort buys: up to 15 points for 16x the tokens

Moving from low to max effort adds 10 to 15 points on the Intelligence Index from Artificial Analysis, but costs eight to sixteen times as many tokens. Opus 5 rises from 51 to 61, Sol from 49 to 59, and Terra from 40 to 55 [9] [10] [11].

That index, v4.1, combines nine evaluations covering terminal work, real-world task completion, GPQA Diamond and long-context reasoning. It weights agent tasks at 34%, coding and scientific reasoning at 24% each, and general reasoning at 18% [12]. More importantly for this comparison, Artificial Analysis ran the same suite at every effort setting, so the changes form a consistent series.

• Opus 5
• Sol
• Terra

Show the data as a table

Effort	Opus 5	Sol	Terra
low	51	49	40
medium	56	54	46
high	59	56	49
xhigh	60	58	52
max	61	59	55

Figure 2 keeps the models in the same order at every setting: Opus 5 leads, Sol stays about two points behind, and Terra remains third. What changes near the top is the value of another step. Opus 5 gains only 0.62 unrounded points from xhigh to max, even as output use rises from 76 million to 100 million tokens [9].

The full token totals show the same trade-off. Opus 5 goes from 12 million output tokens at low to 100 million at max, Sol from 6.6 million to 70 million, and Terra from 5.9 million to 96 million [9] [10] [11]. Terra’s final step is especially costly because moving from xhigh to max almost triples its output for three index points [11].

Fable 5 cannot be traced across the same curve because Artificial Analysis publishes only its max result: 60 points from 87 million output tokens, one rounded point below Opus 5 [13].

Both Anthropic models fell back to Opus 4.8 when their safety layer stopped a prompt [13] [14]. With that caveat established, Artificial Analysis calculates a cost of $2.03 per index task for Opus 5 and $2.75 for Fable 5 [14].

Who wins with the same agent? Opus 5 and Sol tie

Datacurve’s DeepSWE produces no clear winner when all 18 models use the same mini-swe-agent. At max effort, Opus 5 solves 74% ± 4 of tasks and Sol solves 73% ± 3 [1]. Their confidence intervals overlap, so the one-point gap is not significant. Fable 5 and Terra follow at the same 70% [1].

I give DeepSWE the most weight precisely because the shared harness removes one major source of variation. Its 113 tasks cover 91 repositories and five languages, and the authors wrote them from scratch so the models could not have seen solutions during training [1]. Hand-written verifiers test behavior rather than implementation details. The prompts are about half as long as SWE-bench Pro’s, yet the solutions require 5.5 times more code [1].

Measurement	Opus 5	Sol	Fable 5	Terra
Tasks solved, %	74	73	70	70
Cost per task, $	11.84	8.39	21.63	4.95 (Best value in this row)
Output tokens, K	118	60 (Best value in this row)	119	72
Agent steps	99	61 (Best value in this row)	88	76

Figure 3 leaves the solved row unmarked because 74% ± 4 and 73% ± 3 are statistically tied, just as Datacurve’s warning about overlapping confidence intervals suggests [1]. The clearer difference is how each model gets there. Sol uses 60,000 output tokens and 61 steps, compared with Opus 5’s 118,000 tokens and 99 steps, which lowers its cost per attempted task from $11.84 to $8.39 [1].

The cost gap is larger still between Terra and Fable 5. Both solve 70% of the tasks, but Terra costs $4.95 per attempt while Fable costs $21.63 [1]. Dividing those figures by the solve rate gives an optimistic cost per solved task of $7.07 for Terra and $30.90 for Fable. I used the same method in my Kimi K3 analysis.

tasks solved at max effort

70%

level with Fable 5 max

measured, per attempted task

$4.95

Fable 5 max: $21.63

derived, per solved task

$7.07

cheapest of the four at max

Even within one model, max effort is not always the best value. Opus 5 at high scores 73% ± 2 for $6.08 per task, effectively matching Sol max for less money, while Sol at xhigh reaches 71% ± 1 for $4.70 [1]. Fable makes the trade-off plainer: xhigh and max both solve 70%, but xhigh costs $13.41 instead of $21.63 [1]. Datacurve’s cost-versus-score comparison identifies Opus 5 max, Fable 5 high and Sol medium as efficient configurations [1].

Which model writes the most mergeable patch?

Fable 5 xhigh and Opus 5 medium are effectively tied at the top of Cognition’s FrontierCode, which asks whether maintainers would accept a patch [15]. The more useful result is that Opus peaks at medium effort, not max.

The benchmark uses tasks written by maintainers of the repositories under test. Runs score zero if they consult a source containing the solution, and the grader checks correctness, test quality, scope and codebase conventions.

The exact best scores are 53.5 for Fable 5 xhigh, 53.4 for Opus 5 medium, 47.5 for Sol max and 41.3 for Terra max [15].

• Fable 5
• Opus 5
• Sol

Show the data as a table

Effort	Fable 5	Opus 5	Sol
low	48.0	41.9	35.4
medium	49.8	53.4	39.9
high	52.7	48.0	45.1
xhigh	53.5	43.6	46.8
max	51.6	48.0	47.5

Figure 5 shows three different responses to more effort. Sol improves steadily from 35.4 at low to 47.5 at max [15], whereas Fable 5 peaks at 53.5 on xhigh and loses 1.9 points at max. Opus 5 is less predictable: it reaches 53.4 on medium, drops to 43.6 on xhigh, then recovers only to 48.0 on max [15]. Anthropic’s system card shows the same curve and confirms medium as Opus 5’s best setting on the main set [16].

FrontierCode’s grading helps explain why more effort can hurt. The benchmark rewards staying within scope [15] and checks for prohibited implementation patterns [16], while higher effort can lead a model to explore more files and broaden the change. That behavior may help with open research, but it works against a narrowly scoped pull request. I saw the same tendency when Opus 5 ran my week.

Fable’s 0.1-point lead over Opus 5 medium does not establish a real winner because each configuration has only five runs and both pass 58.9% of tasks [15].

Their costs are less ambiguous: Fable xhigh costs $13.09 per rollout, compared with $4.30 for Opus medium. Sol’s best score of 47.5 also remains below Fable’s lowest, 48.0 [15].

Claude therefore leads this patch-quality benchmark. The important boundary is that Claude ran in Claude Code while GPT-5.6 ran in Codex [15].

SWE Atlas splits the work between Sol and Fable 5

In Scale Labs’ SWE Atlas runs, Sol leads codebase comprehension, while Fable 5 leads test writing and refactoring [17] [19] [20]. A separate Artificial Analysis run reverses the comprehension order, which shows how strongly the agent setup can affect the result [18].

Repository comprehension changes with the agent setup

Codebase QnA places a real repository in a container and asks the agent 124 questions that require running the software and tracing its execution. The score leans toward architecture at 35% and root-cause analysis at 30% [17].

Scale Labs reports 46.0 ± 5.0 for Sol xhigh in Codex and 39.0 ± 5.0 for Fable 5 in Claude Code, although both models lost points after security filters made them refuse harmless questions [17]. Opus 5 and Terra have no results, and the previous-generation Opus 4.8 leads the board at 57.26 [17].

Artificial Analysis gets a different order from the same 124 questions. In its Coding Agent Index, Opus 5 xhigh scores 54.8, Fable 5 scores 48.9 and Sol max scores 43.3 [18]. Because the task set stayed fixed while the runs and harnesses changed, the reversal shows how much the surrounding system can affect a leaderboard.

Fable 5 leads test writing and refactoring

Scale Labs’ Test Writing benchmark uses 90 tasks to check whether an agent understands the behavior it is testing. Each generated test must pass against the real implementation, then fail after the relevant code is deliberately changed, so empty tests earn nothing [19]. Fable 5 scores 55.6 ± 5.8 against Sol’s 45.9 ± 6.0, but the overlapping intervals leave both models ranked first [19].

• Fable 5
• Sol

Show the data as a table

Benchmark	Fable 5	Sol
Codebase QnA	39.0	46.0
Test writing	55.6	45.9

Figure 6 gives each model a distinct strength in Scale’s runs: Sol answers more questions about an unfamiliar repository, while Fable writes better tests for one.

Refactoring is Fable’s clearest published win at 54.76 ± 6.76. The 70 tasks cover decomposition, interface changes, extraction and relocation, while graders check for regressions, broken call sites, circular dependencies and dead code [20].

Opus 5, Sol and Terra have no result on this benchmark; the best listed GPT model is the previous-generation GPT-5.5 at 44.79 [20]. Those missing rows limit the comparison, rather than counting as zeroes.

Which model handles open-ended agent tasks best? Opus 5

Opus 5 has the strongest overall case on open-ended agent tasks. Artificial Analysis cannot separate Sol, Opus 5 and Terra confidently on Terminal-Bench 2.1 [21], but Anthropic’s published results put Opus ahead on FrontierBench, OSWorld and AutomationBench [16].

Terminal-Bench leaves the leading three statistically close

Artificial Analysis reports 89.5% for Sol xhigh and 89.1% for Opus 5 max, while Sol max, Terra max and Opus 5 xhigh all score 88.0% [21].

The benchmark asks whether a model can use a shell to reach a verified result. With 89 tasks and three repeats, one changed trial moves the score by about 0.4 points [12]. Fable 5 is further back at 84.6% [21].

Show the data as a table

Configuration	Value
Sol xhigh	89.5%
Opus 5 max	89.1%
Terra max	88.0%
Fable 5 max	84.6%

Figure 7 therefore supports no clear winner between Sol and Opus 5, while Terra stays close despite its lower input price of $2.50 per million tokens [2] [21]. OpenAI’s launch table reports slightly different scores of 88.8% for Sol, 87.4% for Terra and 83.1% for Fable 5 [7]. The shift between OpenAI’s results and the independent run is roughly as large as the gaps between the leading models, which is another reason not to overread their order.

Harder open-ended work favors Opus 5

The same team designed FrontierBench v0.1 around 74 harder tasks in computational biology, physics simulation, CAD, formal proofs and GPU performance [16]. Anthropic ran the models with mini-swe-agent and found a clearer Opus 5 lead: 44.4% at xhigh, 43% at max and 39% at high, with high using 19% fewer output tokens than max. Sol scores 37.5% and Fable 5 scores 33.7%, both at max [16]. For these open-ended problems, Opus 5 has the strongest published result.

Two other benchmarks extend that lead to long, tool-heavy tasks. In OSWorld 2.0, where a model controls a live Ubuntu VM by mouse and keyboard for up to 500 actions per task, Opus 5 scores 70.6%, Fable 5 scores 66.1% and Sol scores 62.6%. Anthropic takes Sol’s result from OpenAI’s release post [16].

Zapier’s AutomationBench instead simulates a company with 47 apps and several business policies. Opus 5 max scores 26.0% and medium scores 24% at $0.89 per task, compared with 18.1% for Sol and 17.4% for Fable [16]. The top AutomationBench row has since changed, and I explain why the leading AutomationBench score belongs to two models.

Opus 5 leads both environments, but the small gap between medium and max again shows that the highest effort is not always necessary.

How to read a 2026 leaderboard: compare setups, not names

A leaderboard measures a complete setup: one model, one agent harness, one effort setting and one grader. Changing any part of that setup can move the score by more than the gap between first and fourth place, so I use five checks before comparing results.

The habit	The exhibit in this comparison
Ask which harness ran	SWE-bench Pro reads Fable 80.0, Opus 5 79.2, Sol 64.6, each vendor’s own agent [16]; DeepSWE’s shared agent compresses the same models into four points [1]
Respect the error bars	74 ± 4 against 73 ± 3 [1]; three exact ties at 88.0% [21]; a 0.1-point lead across five runs [15]
Notice saturation	Opus 5 scores 96.0% on SWE-bench Verified [16]; a board with no headroom ranks nothing
Check the version	OpenAI’s launch page cites Coding Agent Index v1.1 with Sol at 80 [7]; the live v1.3 scores Sol max at 67 [18]
Missing is not zero	Opus 5 has no row on any of Scale’s three boards [17] [19] [20]; Terra barely exists outside DeepSWE and FrontierCode

Version numbers deserve particular attention because a familiar benchmark name can hide a changed test. Coding Agent Index scores the same model family at 80 on v1.1 and 67 on v1.3, but those figures are not comparable. DeepSWE likewise changed GPT-5.5 xhigh from 70% on v1 to 67% on v1.1 [1]. Without the version, the result is incomplete. A later example is the Muse Spark 1.3 rescore on the Artificial Analysis index.

Even primary sources can leave important details unresolved. The Opus 5 system card labels Sol’s FrontierBench run as Codex in its summary table, but the methods section says mini-swe-agent on the same infrastructure and never explains the conflict [16]. Its SWE-bench Multimodal result, 59.4 for Opus 5 against 54.1 for Fable, also comes from an internal harness modified from the public version [16].

Scale introduces further variation by giving newer mini-swe-agent runs 500 steps instead of the previous 250, and reports that leading models perform better in their native harnesses [17]. None of those choices makes the work dishonest, but together they show that many rows measure different systems. My Kimi K3 and GLM-5.2 analysis reached the same conclusion from other data: the capabilities are real, but many scores are not directly comparable.

Three missing results could change these recommendations: Opus 5 on Scale’s benchmarks, a public effort series for Fable 5, and tests of each model in the other vendor’s harness. Until they exist, I would choose for the task rather than buy the highest rank. The most expensive model is rarely the best at everything, Terra stays close on several benchmarks, and turning effort down can improve both cost and patch quality. OpenAI has since shipped a successor to Sol, and my comparison of GPT-6 Astra with Sol 5.6 continues from here.