Hands-on Test of GPT 6.1 Sol and Claude Sonnet 5.5: 3.5x Cost Difference Across Three Tasks

On September 28, Anthropic launched Claude Sonnet 5.5, and on September 29, OpenAI launched GPT 6.1 Sol. Coincidentally, the two models have exactly the same API list price on OpenRouter: $2 per million input tokens and $10 per million output tokens. Leiphone.com Both companies have given highly similar directions: compress the capabilities of near-top-tier models into a faster, cheaper product layer suitable for Agents and professional work. Sonnet 5.5 emphasizes completing well-defined daily tasks in fewer steps; GPT 6.1 Sol targets complex coding, computer operation, and professional work. We put both models into the same workbench, called them through the same channel provider, and asked the models to consecutively complete three tasks: 3D escape room development, latest policy investigation, and Nobel Prize chemistry popular science. In the end, GPT 6.1 Sol cost $38.5, while Claude Sonnet 5.5 used only $10.9, a difference of about 3.5 times. The difference in speed is even more interesting. When developing 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 only 5 minutes, while Sol took nearly an hour. In addition to the final deliverables, we also compared end-to-end time consumption, tool calls, original file archiving, runtime verification, and actual billing. The results did not form a neat win-loss line. GPT 6.1 Sol was faster in escape room development and established a deeper evidence chain and release loop in the two research tasks; Claude Sonnet 5.5 delivered usable results through shorter paths, was several times faster in both research tasks, and its total cost for the three tasks was only about 28% of Sol's. This is a typical kind of "jagged intelligence": Sol is a deep-work workaholic willing to keep verifying, while Sonnet 5.5 is more like an efficient daily vanguard. Through detailed analysis, we found that the model can only determine the upper limit of capability; the toolchain, verification method, and stopping strategy are the core determinants of what we ultimately get and how much we pay for it. These two products also reflect that the industry position of mid-tier models is rising: they are beginning to become the main execution layer for enterprise AI, taking on most complex and high-frequency production tasks. Once model capabilities meet the threshold for use, stable delivery and unit task cost will determine the industrial value of a model more than leading a single benchmark. 01

Comparative hands-on test: 3 real tasks to test the 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. We used Claude Code Cli as the testing platform and called the two models from the OpenRouter channel provider to complete three complex tasks respectively. Both models can search the web, read and write files, run commands, and call a browser.

Test task 1: 3D escape room game development

Use Three.js to make a first-person 3D escape room: 3 rooms, 6 interactive objects.

Hard constraints:

1. There must be a [chain dependency chain] among the objects; it cannot be six parallel puzzles: only after getting A can you use B, only by combining B and C can you open D, and only after D is opened can you enter the 3rd room... Please draw this dependency chain clearly for me to see;

2. Every object click must have clear feedback (if there is feedback, but it is not yet usable, also give a prompt explaining "why it cannot be used now" instead of silently not responding);

After finishing: walk through the clearance process yourself from beginning to end, and write out the operation and feedback for each step.

English text is being prepared.

Test task 2: Latest policy investigation. This question specifically tests time-sensitive information and evidence discipline. Executive orders, international joint statements, export rules, and scientific research security policies are mixed together, making it easy to write political wording as a technological breakthrough, and also easy to expand a search snippet into an established fact.

Please complete an investigation report for Chinese technical practitioners based on US Executive Order 14434, the Kyoto Vision, and related primary materials:

1. Explain what "AI renamed SI" changes and does not change at the legal, technical, and industrial levels respectively;

2. Establish a "claim—original evidence—credibility—possible misreading" table;

3. Analyze the three practical impacts it may have on Chinese AI companies, open-source communities, and scientific research cooperation;

4. List five possible but inaccurate Chinese headlines and correct them one by one;

5. All time-sensitive facts must be accompanied by sources, and policy wording must not be treated as a technological breakthrough. From the running process, Sol first wrote the verification method, then ran the deep research process, and finally delivered an investigation report containing 14 files, including a formal report, source and verification records, a methodology note, 5 official original documents, 4 auxiliary extracted texts, and a SHA-256 checksum list. The 5 original documents totaled about 600,000 bytes, the download results were hash-verified, and the 5 groups of core conclusions were each handed to 3 verification paths, leaving 15 judgments in total, followed by one review and revision of the entire report. Overall, the whole job was very meticulous and strictly followed the requirements to complete all tasks. Screenshot of the GPT 6.1 Sol task running process. Screenshot of the delivery note after the GPT 6.1 Sol task execution ended. Unlike GPT 6.1 Sol, Claude Sonnet 5.5 completed a 172-line, about 23,000-byte report in only 5 minutes 11 seconds, calling tools 32 times in the process. The main text structure is complete, likewise providing a 12-item verification table, three China impacts, five misleading headlines, and listing 4 dates worth continuing to track; at the same time, it also caught the difference in wording between the White House and Chinese texts on "whether they agree to the renaming," an angle of great value. Screenshot of the delivery note after the Claude Sonnet 5.5 task execution ended. From the testers' experience, Sonnet is more like quickly drawing a news map, while Sol works harder to patch the boundaries and ensure data authenticity. Overall, when highly time-sensitive news requires quickly establishing an issue map, Sonnet is more likely to grab the first version; when it involves legal boundaries, China policy, and accountable citations, Sol's evidence package is closer to a draft that can be reviewed.

Test task 3: Nobel Prize chemistry popular science

Please read the official announcement and popular science materials of the 2026 Nobel Prize in Chemistry (the materials need to be obtained by yourself), and complete the following deliverables:

1. Explain chirality, enantiomers, nonlinear effects, and asymmetric autocatalysis in a way ordinary readers can understand

2. Explain which step of the problem Kagan's and Soai's contributions each solved, and they must not be conflated

3. Create a single-file HTML interactive demo showing how a tiny initial bias is amplified in the autocatalytic process

4. Clearly state that the interactive demo is a teaching model, not an exact kinetic reproduction of the Soai reaction

5. Output a five-page presentation outline, and verify all numbers, years, and drug-related statements. This question requires the model to simultaneously play the roles of science editor, front-end developer, and fact-checker. The hardest part is maintaining the separation of three layers of content: award facts come from official materials, the chemistry explanation must be friendly to ordinary people, and the interactive webpage can only show mechanistic intuition and must not pretend to be an exact kinetic reproduction of the Soai reaction. From the results, Sol delivered a result containing 42 files, including 7 documents, 1 interactive page, a five-page presentation outline, and 3 verification scripts. This is the file stock left by the model after multiple rounds of repeated execution; among the 19 scraping records, 13 source files were successfully archived and completed size and hash verification, with core sources covering the Nobel announcement, popular science materials, scientific background, and FDA documents. Screen recording of the GPT 6.1 Sol interactive demo interface. Judging from the delivered files, Sol was extremely attentive to detail handling, especially reflected in many details, for example, Sol clearly distinguished R/S absolute configuration, optical rotation direction, and D/L nomenclature, and at the same time split the timeline into five nodes: Frank 1953, Kagan 1986, Soai 1990, 1995, and 2003, avoiding lumping theory, nonlinear transmission, autocatalysis, and tiny-bias amplification into a single discovery. Next, let's look at Sonnet 5.5's performance: Sonnet's delivery is more compact, with 5 files in total: popular explanation, interactive webpage, five-page outline, fact-check table, and README. Among them, the interactive page designed three modes: "one batch of reactions," "multiple batches relay," and "random side selection," and users can adjust the initial ee, mutual inhibition strength, and catalytic selectivity. Setting the mutual inhibition strength to 0 makes the amplification disappear; the random mode can run 60 times at once, letting readers directly see how tiny fluctuations determine the final direction. Screen recording of the Claude Sonnet 5.5 interactive demo interface. Both models correctly distinguished Kagan, Soai, and the teaching model, but their strengths lie in different links. From the perspective of scientific explanation, Sol emphasizes conceptual boundaries, historical nodes, and correction of drug cases, while Sonnet focuses on fact-checking and other aspects; from the perspective of source coverage, Sol focuses more on Nobel and FDA primary materials, while Sonnet focuses on scientific background documents and at the same time only obtained some webpage summary data as materials, etc. Overall, Sonnet's interactive design and numerical self-checks are richer; Sol's source archiving, FDA original-document correction, and real-browser verification form a more complete release loop. The former is suitable for quickly making a popular science prototype that tells a story and can continue to be iterated, while the latter is suitable for formal releases that require reproducible records and acceptance credentials. 02

Same price, different bills: why did speed and cost reverse three times? From their positioning, the two models were originally in a similar product range. OpenAI defines GPT 6.1 Sol as "providing complex work performance close to Astra at lower cost," with key scenarios including complex coding, computer operation, and professional work; the official page gives a context window of 1.05 million tokens and a maximum output of 128,000 tokens. Anthropic, meanwhile, positions Claude Sonnet 5.5 as a faster, lower-cost model than Opus 5.5, aimed mainly at everyday tasks, bug fixing, and document, slide, spreadsheet, and design work. The AA leaderboard shows that both models are in the top range. The time the two models spent executing tasks is shown in the figure below. In the escape room development, GPT 6.1 Sol took 33 minutes 37 seconds and Claude Sonnet 5.5 took 74 minutes 32 seconds, the latter about 2.2 times the former; for the policy investigation, Sonnet took only 5 minutes 11 seconds, while Sol took 54 minutes 50 seconds, about 10.6 times Sonnet's time. Finally, in the chemistry task, Sol took 65 minutes 41 seconds and Sonnet 12 minutes 54 seconds, a nominal difference of about 5.1 times. A comparison chart of the two models' terminal time statistics on the three test tasks shows through analysis that the time differences come from the two models' different action paths. In the escape room development, Sol completed the implementation and actual playthrough faster; Sonnet reworked multiple times on the Playwright version, browser dependencies, and low-frame-rate environment, and the automated walkthrough lengthened the entire round. In the policy investigation and chemistry task, it was exactly the reverse: Sonnet quickly completed retrieval and the report framework, while Sol invested more time in original-document archiving, hash verification, multi-path verification, and full-text review. The former prioritizes producing an editable first draft, while the latter includes traceability within the delivery scope. This also shows that an Agent's wall-clock time is simultaneously affected by model strategy, tool availability, environment failures, and acceptance depth, and is hard to compress into a fixed "speed ranking." Finally, in terms of usage cost, the two have the same API price on OpenRouter: $2 per million input tokens, $10 for output, and $2.5 for cache writes. As the billing panels for the three tasks below show, GPT 6.1 Sol cost a total of $38.5, while Claude Sonnet 5.5 cost $10.9, the former about 3.53 times the latter. The two models' expense summary statistics across the three tasks have the same list price, yet the total bills still differ by 3.5 times, mainly because of the Agent's action paths. Sol launched multi-path verification in the policy investigation, saved large numbers of original documents and ran a real browser in the chemistry task, and repeatedly reran the full workflow in the escape room development; more tool calls, longer contexts, and more repair rounds all go into the bill. Sonnet more often chooses the shortest deliverable path, especially evident in the policy question. 03

Next stop: who can carry a task through to the finish line? This round of hands-on testing reflects a clear change: frontier large models are moving from content generators into task execution systems, where models search for materials, read original documents, write programs, call browsers, fix errors, and organize deliverables. The yardstick for measuring models has expanded accordingly: answer quality still matters, but whether the process can be reviewed, whether failures are visible, and whether costs are controllable also determine whether it can enter real work. From an industry-layering 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 capability ceiling, small models handle high-frequency simple tasks, and mid-tier models take on the largest number of complex production tasks. What enterprises actually buy is stable, verifiable, and cost-predictable task success; the token unit price is only one of the costs. GPT 6.1 Sol demonstrates its potential to take on "research engineering" and "quality assurance" work, delivering raw materials, hashes, test scripts, browser logs, and failure records together, and also adding a review layer for high-risk conclusions. This capability raises the ceiling for autonomous execution of complex tasks, letting models extend from producing content to verifying content. At the same time, the corresponding price is longer action chains and higher resource consumption: this round's total cost was $38.5, the policy question took nearly an hour, and multi-path review cannot naturally add independent evidence. Claude Sonnet 5.5 demonstrates another equally important development path: compressing strong reasoning and production capabilities into a shorter, cheaper execution path, quickly building a complete report and a deliverable scientific demonstration, and proactively declaring boundaries such as missing original documents and unverified browsers. In addition, the total cost of $10.9 is about 28% of Sol's, already falling into a cost range that is easier to use for everyday work. However, limitations also exist: the short path sometimes stops at a usable prototype, and visual acceptance and peripheral sources still require manual handling. The changes brought by the two models therefore fall in different directions. Sol shows "to what extent a model can verify itself"; Sonnet 5.5 lowers the time and price threshold for high-quality Agents to enter high-frequency tasks. Simply put, Sonnet 5.5 is better suited as a high-performance "daily workhorse" model, with better overall scores, coding tasks, and generation speed; 6.1 Sol is better suited as a deep Agent for unit tasks, suitable for finely polishing specific task details. In future work, different models can take on different roles within the same workflow, with a fast model exploring and drafting, and then a cautious model checking key facts and execution paths. At a deeper level, AI competition is expanding from a model's single-point capability to whole-system capability: the model affects the reasoning ceiling, tools and the runtime environment determine which actions it can execute, the verification mechanism affects the credibility of results, and the stopping strategy affects time and bills. The repeatedly reversing speed and quality advantages across the three questions show that a single benchmark can hardly cover these variables; an Agent's real performance comes from the combined effect of the model, toolchain, task design, and acceptance standards. AI is moving past the stage of "giving a good answer" and beginning to take on the continuous process from understanding a problem to delivering a result. ▼ Recommended Reading ▼

WorkBuddy, Doubao Work, and Qwen Office hands-on test: AI has finished the work, so why are you more tired?

2026-09-18

Seedance 2.5 vs Kling 3.0: one is better at shooting, the other at producing finished footage

2026-09-11