Google's Gemini 4 Argon matched Anthropic's Claude Opus 5.5 on code debugging but delivered inaccurate results on long-context reasoning tasks, according to a head-to-head test published by The New Stack on October 19, 2026. The evaluation tested both models on three challenging tasks designed to mirror categories where Google claimed benchmark advantages when it opened Argon to paid API developers on October 13. While Argon performed at half the cost during a promotional pricing window, accuracy differences emerged that call Google's benchmark claims into question.
In the code review challenge, both models identified all 14 logic errors planted across a 3,500-line Python library without flagging any of the five decoys, and both successfully repaired every bug without breaking existing tests. Argon's review averaged 1 cent per bug discovered versus 2.4 cents for Opus, consuming 53,031 input tokens and 3,964 output tokens compared to Opus's 64,572 and 4,053. The billing specification test revealed a sharp divide: Opus passed all 80 hidden tests and then all 100 after a currency expansion, while Argon passed just 11 of 80 and 20 of 100 due to a single incorrect argument passed to Python's datetime function early in the code. On long-context reasoning, Opus correctly identified 33 services in a dependency failure scenario at 500,000 tokens and 32 at 600,000 tokens with no incorrect entries, while Argon returned 20 correct and 8 wrong at 500,000 tokens, 19 correct and 16 wrong at 600,000 tokens, and 23 correct and 2 wrong at 900,000 tokens.
The tester noted that Argon "got the reasoning wrong across the entire answer" on the dependency mapping exercise, calling it "a big red flag." Argon raced through the long-context test without displaying step-by-step work, using just 243 to 425 output tokens compared to Opus's over 5,000 tokens, most devoted to reasoning chains. The author stated they "wouldn't unquestioningly trust Argon on any task" despite its lower cost and competitive performance on code review. Google's promotional pricing sets Argon at $2 per million input tokens and $10 per million output tokens through November 30, then rises to $4 and $20 on December 1 to match Opus 5.5's standard rates.
The report concludes that while Google is working hard to catch up to Anthropic, it hasn't succeeded yet, citing Artificial Analysis's independent ranking that places Opus 5.5 at position 58 ahead of Argon's 53. The billing test's single-line error suggests Argon can produce structurally sound code that fails on implementation details, while the dependency reasoning mistakes point to deeper flaws in multi-hop logic across large contexts. The six-week pricing window ending November 30 gives developers a narrow period to evaluate whether Argon delivers genuine quality parity or simply undercuts competitors on price, but the tester's final assessment warns that benchmarks don't always reflect real-world task performance. For teams considering a switch, the gap between advertised benchmark wins and actual reasoning accuracy may prove more costly than any temporary discount. The industry now faces a familiar question: whether aggressive pricing and selective benchmark marketing can substitute for the methodical accuracy that production workloads demand.