Anthropic released Claude Sonnet 5.5 on September 28, and OpenAI released GPT 6.1 Sol on September 29. Coincidentally, the two models have exactly the same listed API prices on OpenRouter: $2 per million input tokens and $10 per million output tokens. Leiphone(WeChat public account: Leiphone)
The two companies are moving in highly similar directions: compressing the capabilities of near-top-tiermodels into a product tier that is faster, cheaper, and suited to agents and professional work. Sonnet 5.5 emphasizes completing everyday tasks with clearly defined boundaries in fewer steps; GPT 6.1 Sol targets complex coding, computer use, and professional work.
We placed both models on the same workbench, invoked them through the same channel provider, and tasked them with completing three jobs in succession: developing a 3D escape room, investigating the latest policies, and creating a popular-science explainer on the Nobel Prize in Chemistry. In the end, GPT 6.1 Sol cost $38.5, while Claude Sonnet 5.5 cost only $10.9—a gap of roughly 3.5×.
The differences in speed are even more interesting. When building the 3D escape room, Sol took 33 minutes, while Sonnet took 74 minutes; for the policy investigation, the situation completely reversed, with Sonnet completing the report in just 5 minutes while Sol took nearly an hour.
In addition to the final deliverables, we compared end-to-end completion time, tool calls, archiving of original materials, runtime validation, and actual billing charges.
The results did not produce a clear-cut winner: GPT 6.1 Sol was faster at escape-room development and built deeper evidence chains and closed-loop publishing workflows for the two research tasks; Claude Sonnet 5.5, meanwhile, delivered usable results via a shorter path, was several times faster on both research tasks, and cost only about 28% of Sol’s total across all three tasks.
This is a typical case of “jagged intelligence”: Sol is a workaholic for deep work, willing to keep verifying, while Sonnet 5.5 is more like an efficient everyday trailblazer. Detailed analysis shows that models determine only the upper bound of capability; the toolchain, validation methods, and stopping strategy are the core factors determining what we ultimately get and what it costs us.
These two products also reflect the rising industry position of mid-tier models: they are beginning to become the primary execution layer of enterprise AI, taking on most complex yet high-frequency production tasks.Once a model’s capabilities meet the threshold for practical use, reliable delivery and cost per task will have a greater bearing on its industrial value than topping individual leaderboards..01
Comparative Hands-On Test: Three Real-World Tasks to Evaluate Two Models
This article selects two recently released top-tier models for comparative testing: GPT 6.1 Sol and Claude Sonnet 5.5 from OpenAI and Anthropic, respectively. Using Claude Code CLI as the test platform, we called both models through the OpenRouter provider to complete three complex tasks. Both models can search the web, read and write files, run commands, and invoke a browser.Test Task 1: Developing a 3D Escape Room Game
Build a first-person 3D escape room using Three.js: 3 rooms and 6 interactive objects.
Hard Constraints:
TAG150@@1. The objects must form a chain of dependencies and cannot be six independent puzzles: you must obtain A before you can use B, B must be combined with C to open D, and D must be opened before you can enter the third room… Please explicitly draw out this dependency chain for me;
2. Clicking each object must produce clear feedback (even if the object cannot yet be used, it must explain “why it cannot be used right now” rather than remaining silently unresponsive);
3. There must be a “Current Objective” prompt so that the player cannot get stuck;
4. The entire game must be completable within 5 minutes;
5. All files and intermediate files must be saved in the current folder.
TAG180@@When you are done: play through the entire game yourself from start to finish and write out the action and feedback for every step.
If you cannot complete the game yourself, directly tell me which step you got stuck on.
The model must implement both 3 rooms and 6 items and connect them into a dependency graph; it must provide feedback when players click randomly, click too early, or act from too far away; and finally, the model itself has to play through it from beginning to end.Let’s see how the two models perform:First up is 6.1 Sol. The 《Peripheral Vision / AFTERLIGHT》 it created features a fully realized retro control-room aesthetic: green wainscoting, checkered flooring, 3D signage, an inventory bar, a sound toggle, and a persistent objective prompt have all been incorporated into the finished product. The six items are arranged into a dependency chain in which two branches converge: Fuse A restores power to Distribution Box B, while Access Card C, together with the powered state, unlocks Terminal D; only then can the player enter the third room, obtain Key E, and open Exit F.The gameplay experience of the 3D escape-room game built by GPT 6.1 SolUnlike Sol, Sonnet 5.5 uses a gray cell-and-electrical-room aesthetic, with fewer visual decorations, but the puzzle chain remains complete. Players must first retrieve a screwdriver from under the mattress, then remove the air vent to obtain the administrator’s key; the key opens both the cell door and the desk drawer, while the access card in the drawer unlocks the tool cabinet and yields the battery. Finally, the “access card + battery” together open the electronic exit door. The glowing doorway at the far end of the third room serves only to trigger the results screen and is not sneakily counted among the six interactive objects.The gameplay experience of the 3D escape-room game built by Claude Sonnet 5.5Based on the evaluators’ impressions, Sol feels more like a dry, play-by-play diary of an assignment being handed in, while Sonnet clearly creates a sense of suspense for the escape-room experience. Every step in Sol is explicitly marked, and the access card’s design looks more like a floating object that defies the laws of physics. By contrast, both the texture of the key and the keyhole on the door in Sonnet convey a sense of realism.Test Task 2: Latest Policy Research
Based on U.S. Executive Order 14434, the “Kyoto Vision,” and related primary sources, write an investigative report aimed at Chinese technology practitioners:
1. Explain what the renaming of “AI” to “SI” does—and does not—change at the legal, technical, and industrial levels;
2. Create a “Claim—Primary Evidence—Confidence Level—Possible Misreading” table;
3. Analyze three practical impacts it may have on China’s AI companies, the open-source community, and research cooperation;
4. List five Chinese headlines that might appear but would be inaccurate, and correct each one;
5.All time-sensitive facts must include sources, and policy language must not be presented as a technological breakthrough.
Judging from the execution process, Sol first developed its verification methodology, then ran a deep-research workflow and ultimately delivered a research report containing 14 files, including the formal report, source and verification records, a methodology description, five official source documents, four auxiliary extracted texts, and a SHA-256 checksum manifest. The five original files totaled approximately 600,000 bytes, and the downloaded results were hash-verified. It also assigned the five sets of core findings to three separate verification tracks, producing 15 assessments in total, before conducting a review and revision of the entire report. Overall, the work was extremely thorough and strictly fulfilled all requirements.
Screenshot of the GPT 6.1 Sol task execution process
Screenshot of the delivery instructions after the GPT 6.1 Sol task was completedUnlike GPT 6.1 Sol, Claude Sonnet 5.5 completed a 172-line report of approximately 23,000 bytes in just 5 minutes 11 seconds, using 32 tool calls throughout the process. Its main body was structurally complete and likewise included a 12-entry verification table, three implications for China, five misleading headlines, and four dates worth continuing to monitor. It also identified the difference in wording between the White House and the Chinese side’s texts regarding “whether the name change was agreed to,” an especially valuable angle. 
Screenshot of the delivery instructions after the Claude Sonnet 5.5 task was completedFrom the testers’ perspective, Sonnet was more like someone rapidly sketching a map of the news, whereas Sol was more focused on patching boundaries to ensure the authenticity of the data. Overall, when highly time-sensitive news requires a rapid issue map, Sonnet is more likely to produce the first draft quickly; when legal boundaries, China-related policies, and accountable citations are involved, Sol’s evidence package more closely resembles working documentation that can be independently reviewed. Test Task 3: Popular-Science Explanation of the Nobel Prize in Chemistry
Please read the official announcement and popular-science materials for the 2026 Nobel Prize in Chemistry (the materials must be obtained independently), and complete the following deliverables:
1. Explain chirality, enantiomers, nonlinear effects, and asymmetric autocatalysis in terms that ordinary readers can understand
2. Explain which step each of Kagan’s and Soai’s contributions addressed; the two must not be conflated
3. Create a single-file interactive HTML demonstration showing how a minute initial deviation can be amplified during an autocatalytic process
4.Explicitly state that the interactive demo is a teaching model, not an exact kinetic reproduction of the Soai reaction
5. Produce a five-slide presentation outline and verify all figures, years, and drug-related statements.
The task requires the model to play the roles of science editor, front-end developer, and fact-checker simultaneously. The most difficult part is keeping three layers of content separate: award facts must come from official materials, the chemistry explanations must be accessible to nonexperts, and the interactive webpage may only convey mechanistic intuition rather than pass itself off as an exact kinetic reproduction of the Soai reaction.In the end, Sol delivered 42 files, including seven documents, one interactive page, a five-slide presentation outline, and three validation scripts. This was the file inventory remaining after the model had undergone multiple rounds of repeated execution. Of 19 retrieval records, 13 source files were successfully archived and passed file-size and hash checks. The core sources covered the Nobel Prize announcement, educational materials, scientific background, and FDA documents.GPT 6.1 Sol interactive demo interface screen recordingJudging from the delivered files, Sol was extremely attentive to detail in many places. For example, Sol clearly distinguished between R/S absolute configuration, direction of optical rotation, and D/L nomenclature. It also divided the timeline into five milestones—Frank 1953, Kagan 1986, Soai 1990, 1995, and 2003—avoiding the blending of theory, nonlinear transmission, autocatalysis, and amplification of tiny deviations into a single discovery.Next, let us look at Sonnet 5.5’s performance: Sonnet’s delivery was more compact, comprising five files in total: an accessible explanation, an interactive webpage, a five-slide outline, a fact-checking table, and a README. The interactive page offered three modes—“single-batch reaction,” “multi-batch relay,” and “random side selection”—allowing users to adjust the initial ee, mutual inhibition strength, and catalytic selectivity. Setting mutual inhibition strength to 0 eliminated the amplification accordingly; random mode could run 60 iterations at a time, allowing readers to see directly how tiny fluctuations determine the final direction.Claude Sonnet 5.5 interactive demo interface screen recordingBoth models correctly distinguished Kagan, Soai, and the teaching model, but their strengths lay in different areas. In terms of scientific explanation, Sol placed greater emphasis on conceptual boundaries, historical milestones, and correcting drug-related examples, whereas Sonnet focused more on fact-checking and related aspects. In terms of the breadth of information sources, Sol focused more on primary materials from the Nobel Prize and FDA, whereas Sonnet focused more on scientific-background documents while retrieving only some webpage summary data as source material, among other things.Overall, Sonnet offered richer interactive design and numerical self-checks; Sol’s source archiving, corrections based on original FDA documents, and real-browser verification created a more complete release loop. The former is well suited to quickly producing a narrative-driven popular-science prototype that can continue to be iterated upon, while the latter is better suited to formal releases requiring reproducibility records and acceptance evidence.02
Same Price, Different Bills: Why Did Speed and Cost Reverse Three Times?
In terms of positioning, the two models were originally in similar product tiers. OpenAI defined GPT 6.1 Sol as providing “complex-work performance close to Astra at a lower cost,” with key use cases including complex coding, computer operation, and professional work. The official page listed a context window of 1.05 million tokens and a maximum output of 128,000 tokens.Anthropic positions Claude Sonnet 5.5 as a faster, lower-cost model than Opus 5.5, primarily targeting everyday tasks, bug fixes, and document, slide, spreadsheet, and design work.
The AA leaderboard shows that both models fall within the top tier.The time spent by the two models executing the tasks is shown below. In escape-room development, GPT 6.1 Sol took 33 minutes and 37 seconds, while Claude Sonnet 5.5 took 74 minutes and 32 seconds, with the latter taking approximately 2.2 times as long as the former. In the policy research task, Sonnet took only 5 minutes and 11 seconds, while Sol took 54 minutes and 50 seconds—approximately 10.6 times as long as Sonnet. Finally, in the chemistry task, Sol took 65 minutes and 41 seconds and Sonnet took 12 minutes and 54 seconds, a recorded difference of approximately 5.1 times.
A comparison chart of terminal-time statistics for the two models across three test tasksAnalysis reveals that the time difference stems from the models’ different action paths. In escape-room development, Sol completed the implementation and actual playthrough faster. Sonnet repeatedly had to redo work because of the Playwright version, browser dependencies, and the low-frame-rate environment, while its automated walkthroughs extended the entire run. The opposite occurred in the policy research and chemistry tasks: Sonnet quickly completed the research and report framework, while Sol devoted more time to archiving original materials, hash verification, multi-pronged verification, and full-text review. The former prioritized producing an editable first draft, while the latter also incorporated traceability into the deliverables.This also shows that an Agent’s wall-clock time is jointly affected by the model’s strategy, tool availability, environmental failures, and the depth of acceptance testing, making it difficult to reduce to a fixed “speed ranking.”Finally, in terms of usage costs, the two models have identical API pricing on OpenRouter: $2 per million input tokens, $10 per million output tokens, and $2.50 per million tokens for cache writes. As the billing panels for the three tasks below show, GPT 6.1 Sol cost $38.50 in total, while Claude Sonnet 5.5 cost $10.90, with the former spending approximately 3.53 times as much as the latter.
TAG465@@A summary of the two models’ spending across the three tasksDespite having the same list price, their total bills differed by 3.5 times, primarily because of the Agents’ different action paths. In policy research, Sol initiated multi-pronged verification; in the chemistry task, it saved a large number of original files and used a real browser; and in escape-room development, it repeatedly reran the entire workflow. More tool calls, longer contexts, and additional repair rounds are all billed. Sonnet more often chose the shortest path to a deliverable, especially evident in the policy task.TAG474@@TAG476@@03
TAG480@@
TAG483Next, who can bring the task across the finish line?TAG489TAG492TAG494@@This round of real-world testing reveals a clear shift: frontier large language models are moving from content generators into task-execution systems, with models searching for information, reading original files, writing programs, invoking browsers, fixing errors, and organizing deliverables. The yardstick for evaluating models is expanding accordingly. Answer quality remains important, but whether the process can be reviewed, failures are visible, and costs remain controllable also determines whether a model can enter real-world work.
From an industry-tier perspective, Sol and Sonnet represent the two companies’ mid-tier models and are becoming the default working layer of enterprise AI systems:flagship models explore the upper limits of capability, small models handle frequent, straightforward tasks, and mid-tier models take on the largest volume of complex production tasks. What enterprises are really buying is stable, verifiable task success at a predictable cost; the unit price of tokens is only one of those costs.GPT 6.1 Sol demonstrates the potential to take on “research engineering” and “quality assurance” work, delivering source materials, hashes, test scripts, browser logs, and failure records together while adding a review layer for high-risk conclusions. This capability raises the ceiling for autonomous execution of complex tasks, allowing the model to extend from producing content to verifying it. The corresponding costs are longer action chains and greater resource consumption: this run cost a total of $38.50, the policy question took nearly an hour, and multi-path review cannot inherently produce additional independent evidence.Claude Sonnet 5.5 demonstrates another equally important development path: compressing strong reasoning and content-creation capabilities into shorter, less expensive execution paths that can rapidly produce complete reports and deliverable scientific demos while explicitly stating limitations such as missing original documents and unverified browser results. Moreover, its total cost of $10.90 was about 28% of Sol’s, already placing it in a cost range that is easier to use for everyday work. The limitations are equally present: shorter paths sometimes stop at a usable prototype, while visual acceptance and peripheral sources still require human handling.The changes brought by the two models therefore point in different directions. Sol shows “how far a model can verify things on its own,” while Sonnet 5.5 lowers the time and price thresholds for bringing high-quality agents into high-frequency tasks. In short, Sonnet 5.5 is better suited to serve as a high-performance “daily workhorse” model, with better overall scores, coding performance, and generation speed; 6.1 Sol is better suited to serve as a deep agent for individual tasks, refining the details of specific assignments. In future work,different models can take on different roles within the same workflow, with a fast model exploring and drafting before a cautious model checks critical facts and execution paths.At a deeper level, AI competition is expanding from the isolated capabilities of individual models to the capabilities of entire systems.The model affects the upper limit of reasoning, tools and the runtime environment determine which actions it can execute, verification mechanisms affect the credibility of results, and stopping policies affect time and billing. The repeatedly shifting speed and quality advantages across the three questions show that a single benchmark can hardly capture all these variables; an agent’s real performance comes from the combined effects of the model, toolchain, task design, and acceptance criteria.AI is moving beyond the stage of “giving a good answer” and beginning to shoulder the continuous process from understanding a problem to delivering a result.
▼
Recommended Reading
▼
2026-09-18
TAG602@@Seedance 2.5 vs Kling 3.0: One Is Better at Filming, the Other at Producing Polished Videos
2026-09-11
Original article from LeiFengNet. Unauthorized reposting is prohibited. See the Reposting Notice for details.Reposting Notice.
