Claude Opus 5.5 vs GPT-6 Sol: Ten Real Tasks, One Lopsided Scoreboard

September 22, 2026 was a double release day. Anthropic shipped Claude Opus 5.5, and OpenAI followed roughly an hour later with GPT-6 Sol and GPT-6 Luna. The Nate Herk channel put both models to work on ten practical jobs the same day: building websites, editing video, assembling slide decks, browser work, heavy reasoning and a few wildcard tasks, with identical prompts on both sides. The scoreboard read seven wins for Opus 5.5, one for GPT-6 Sol and two rounds voided by the tester's own setup mistakes. The more interesting story sits below the scoreline, because Sol launched at exactly half the price of its predecessor family, and the test shows what that discount buys.

The price table frames the whole day. Claude Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, GPT-6 Sol costs $2 and $10, cached input sits at $0.20 for both. Simon Willison tabulated the rates on release day and called the double launch the opening of a new price war, noting that Sol and Luna each cost half of their direct GPT-5.6 equivalents. The lazy read is that the pricier model must be the better one. The test flips that assumption at a specific point: the price gap is real, and so is the quality gap, and no discount argument removes the second one.

The methodology stays deliberately plain, which is what makes it useful. Both models received the exact same prompt in every one of the ten trials, running through the Claude Code and Codex desktop surfaces, with quality, wall clock time and dollar cost tracked per task. A warm up round came first: two Fireflies call transcripts from Q4 planning, about four hours of material in total, had to be ingested into a wiki while a parallel job searched for the last meeting where one specific topic came up. Sol visibly processed both jobs at once and felt quicker, Opus needed about six minutes and came back with the cleaner result. Sol dated the searched mention to August 17, Opus returned September 14, and September 14 was correct. Speed is not accuracy.

The tester's mental model is worth borrowing. Instead of comparing benchmark curves he asks a simpler question: hand both models a hundred dollars and check which one returns more value for the money. On list prices Sol should win that game by default, since its output tariff is half of the Opus rate. In practice the expensive model was the better investment in seven of the eight scored rounds, and in the two voided rounds the fault sat with the tester, not with either model.

Across all ten workloads Opus 5.5 accumulated 8 hours 40 minutes of runtime and roughly $213 of spend, while GPT-6 Sol finished in 5 hours 51 minutes at $74.46. Scaled out, the premium model ran about 1.5 times slower at three times the cost, and still took seven of the eight scored rounds. The practical translation: Sol is the cheap fleet worker for routine volume, Opus 5.5 is the specialist you reach for when a second look matters. That makes the buying decision a real one again, less about which model is generally better and more about which mix of premium reasoning and discount execution fits a given workload.

The starkest gap appeared in the browser drawing test. Both models got an AI generated reference image and the instruction to recreate it inside Canva with the drawing and painting tools, image generation explicitly not allowed. Opus 5.5 produced a recreation the tester called surprisingly good, better than what GPT-6 Astra had managed on the same exercise weeks earlier, despite running about 30 minutes longer at roughly four times the cost. Sol first tried to shortcut the task by generating an image, and after two corrective prompts still delivered comically poor canvases. The tester read the episode as a regression in browser control from Astra to Sol and scored the round clearly for Claude Code.

The quiet counterweight came in the full course build. Both systems produced nearly identical course content, with the same descriptions, key points and resource links. The differences lived in the surroundings. Opus set only the overall course to draft rather than every single page, which the tester judged the smarter behavior, and it was faster and cheaper on this task. Sol double linked several resources and stacked avoidable redundancy into the result. The expected rule, cheaper worker wins when output matches, did not hold here: with quality equal, time and cost decided, and this time they favored the expensive option. For the channel it also marked a rare browser adjacent win for Claude Code in this matchup.

Early community reactions line up with the video. Willison stresses the pricing front, where Sol and Luna halve the rates of their GPT-5.6 counterparts, and treats that as the actual event of the day. A Reddit thread from release day reports Opus winning every raw benchmark score in its comparison table while Sol lands roughly 100 Elo points below the GPT-5.6 Sol it replaces. The picture is not uniformly bad for OpenAI: Sol stays remarkably fast across routine work and remains under a third of the cost, and the tester sees a legitimate role for it in tightly specified, stepwise workflows. But the discount visibly pays for itself in quality, and the video's summary line makes the point bluntly: the low price now has an explanation.

The tester's verdict fits in one sentence: Opus 5.5 is a major step up from Opus 5, while GPT-6 Sol feels like a step below GPT-5.6 Sol, closer to an expected Luna tier. His recommended architecture follows from that: Opus 5.5 as the orchestrator holding oversight, creativity and judgment, with Sol workers executing precisely specified individual steps, and Astra as the option for rigid, fully predefined flows. That expensive head, cheap hands pattern is already how most economical agent chains get run, and the release day confirmed it rather than overturning it.

The release day exposes two opposing strategies. Anthropic is selling quality at the top of the price range and, on this evidence, delivered a genuine jump. OpenAI is selling price, cutting list rates in half and, per this test, visibly paying for it in execution quality, most clearly on browser tasks. For buyers this is the first real decision in a while rather than a default: teams that need judgment, creative layout work or long running browser control will keep paying the Opus premium, while high volume routine automation gets a fast, cheap worker in Sol. Whether a sharpened Sol follows, or the war escalates through Luna and Terra, is the question the next few weeks will answer.