Gemini 3.7 Flash vs Sonnet 5 vs GPT-5.6 Terra: Real Wins

Gemini 3.7 Flash benchmarks look very different depending on which table you read. Google’s launch post led with five rows; the full model card publishes nineteen, compared across five models — Gemini 3.7 Flash, Gemini 3.6 Flash, Claude Sonnet 5, GPT-5.6 Terra, and Muse Spark 1.2. This post reads all nineteen.

The stakes are practical. Gemini 3.7 Flash shipped on August 13, 2026 at an introductory $0.75 per million input tokens and $3.75 per million output on Google’s API — a rate that undercuts every rival in its own comparison table, matched only by 3.6 Flash on the same introductory window. If the benchmark story holds, that price rewrites routing decisions. If it only holds on the five rows the vendor highlighted, it doesn’t. We covered the launch itself — pricing mechanics, positioning, and the asterisk on “half price” — in the full launch rundown; this post is the benchmark deep-dive.

What follows: the full 19-row table transcribed from the model card, a recomputed win tally, the rows where Sonnet 5 and even 3.6 Flash quietly beat the new model, the blank cell that most coverage repeated without investigating, the independent Artificial Analysis numbers that actually moved this time, and every price labelled by the surface it comes from. Every benchmark figure below is vendor-stated from Google’s card unless noted otherwise.

Google introduced Gemini 3.7 Flash as “our most intelligent workhorse model yet for coding and agents,” three weeks after 3.6 Flash’s launch carried nearly identical positioning. The announcement post and almost every piece of launch coverage we located anchor on the same five benchmarks: FrontierCode, Terminal-bench 2.1, GDM-MRCR, OSWorld-2.0, and Harvey LAB-AA.

The model card is a different document. It compares five models across nineteen rows — and the two extra dimensions matter. The fifth column, Muse Spark 1.2, wins one row outright and is silently dropped by most coverage. And eight of the ten rows 3.7 Flash doesn’t win sit in the fourteen the press cut left out.

None of this means the table is dishonest — publishing rows you lose is more disclosure than most vendors manage. It means the table rewards close reading. The rest of this post is that reading.

The table below transcribes every benchmark row from the Gemini 3.7 Flash model card, all five model columns included. Bold marks the best score in each row. Recomputing the tally across all nineteen rows: 3.7 Flash wins 9, GPT-5.6 Terra wins 6, Sonnet 5 wins 2 outright, 3.6 Flash wins 1 (against its own successor), and Muse Spark 1.2 wins 1. Against the two rivals Google names in its framing, Sonnet 5 takes three rows — GDPVal-AA, Agent’s Last Exam, and BioMysteryBench’s human-solvable subset — though on GDPVal-AA the whole table’s top score belongs to Muse Spark 1.2.

* GDM-MRCR is a Google DeepMind-authored benchmark, self-reported on every Gemini release — read it as a vendor-designed eval, not an independent standard. † Google’s model card lists 48.6% for 3.6 Flash on DeepSWE v1.1; Google’s own announcement prose, published the same day, states 49.0% for the identical metric. We use the card as the table of record and flag the 0.4pp inconsistency rather than silently reconciling it. ‡ Not published in Google’s table. Anthropic reports Sonnet 5’s computer use on a different benchmark variant, OSWorld-Verified — see Section 05. “—” means the model card lists no score for that cell.

Nine wins out of nineteen is a real result, and some of the wins are meaningful. The headline coding claim holds — narrowly: 43.6% on FrontierCode 1.1 against Sonnet 5’s 42.7%, a 0.9-point edge on production code quality. The WebDev Arena Elo of 1588 leads the field, and it is Arena.ai’s human-preference data rather than a private Google metric — though it reaches the card vendor-reported. The clearest daylight shows up on document and multimodal work: 34.0% on GDP.pdf against Sonnet 5’s 28.0%, 85.4% on LVBench long-video understanding against Sonnet 5’s 68.5% — video remains a Gemini strength that the text-first rivals don’t contest — and 53.6% on HLE-Verified, where Sonnet 5 posts a notably weak 31.0%, 22.6 points behind the leader.

Notice the pattern in those three caps: the narrowest win is on an independent benchmark, and the two most decisive wins are on a private task set and a Google-authored eval. That doesn’t make the numbers false — it makes them exactly the kind of numbers you weight down when comparing across vendors, the same discount you’d apply to any lab grading its own homework.

GPT-5.6 Terra is the strongest counterweight in Google’s own data: six rows, concentrated exactly where Google’s framing claims 3.7 Flash leads — agentic coding and computer use. Terra takes Terminal-bench 2.1 (87.4% vs 85.8%), Terminal-bench 3.0 (20.8% vs 14.9%, a hard new benchmark where the whole field scores low), DeepSWE v1.1 (69.6% vs 65.3%), and OSWorld-2.0 (50.2% vs 47.9%). Sonnet 5 takes Agent’s Last Exam outright at 33.3% — seven points clear of 3.7 Flash on multimodal desktop tasks — and edges BioMysteryBench’s human-solvable subset. On GDPVal-AA knowledge work, Sonnet 5’s 1598 Elo beats 3.7 Flash’s 1525 by 73 points, and Muse Spark 1.2 tops the whole row at 1628.

The oddest entries are the two CharXiv rows, where Gemini 3.6 Flash beats its own successor — 85.2% vs 84.5% without tools and 89.4% vs 88.7% with tools. That is genuinely what Google’s table says, and it deserves credit for printing a same-vendor regression rather than trimming the rows. It is also a useful calibration: when a three-week successor loses to its predecessor on chart reasoning, the version bump bought capability in some places by spending it in others — normal for fast-cycle releases, invisible in a five-row press cut.

Two traps in this table will produce confidently wrong takes, and most launch coverage walked past both. The first is the blank Sonnet 5 cell on OSWorld-2.0. It is not a zero, and it is not evidence that Anthropic declined to disclose computer-use performance. Anthropic publishes Sonnet 5’s computer-use score on a different benchmark variant — OSWorld-Verified, where it reportedly scores 81.2% — and disclosed that it changed its OSWorld-Verified methodology between Sonnet 4.6 and Sonnet 5, restating Sonnet 4.6 to 78.5% in the process. Google tested OSWorld-2.0. Different variant, different methodology, different scale: the 81.2% and the 47.9%–50.2% range in Google’s table must never be cross-compared as the same number.

Both traps are instances of the general failure mode we catalogued in our guide to reading vendor benchmark tables: version mismatches and missing cells get read as verdicts when they are artifacts of who tested what, under which harness. Add this card’s own contribution to the genre — the DeepSWE figure for 3.6 Flash that reads 48.6% in the table and 49.0% in the announcement prose published the same day — and the lesson generalizes: even first-party numbers disagree with themselves at the margins, which is precisely why margins under a point should never drive a routing decision.

This is the part of the story that vendor tables can’t settle, and it is where 3.7 Flash earns its most defensible claim. Artificial Analysis scores Gemini 3.7 Flash (high) at 56 on its Intelligence Index — a 4-point improvement over Gemini 3.6 Flash’s 52, with both scores computed at the same index vintage (v4.1.1, a nine-evaluation composite). That matters because of what came before: our July read of 3.6 Flash found a flat Intelligence Index — “cheaper, not smarter.” That pattern breaks here. AA also puts 3.7 Flash on its Intelligence-vs-Time Pareto frontier at 1.7 minutes per task, which it describes as “40% faster than GPT-5.6 Terra (max),” with output throughput around 340 tokens per second.

Two qualifiers keep the independent read honest. First, the effort-ladder asymmetry from Section 02 applies here too: 3.7 Flash’s 56 is the top of a three-rung ladder (51/53/56), while Terra’s 57 comes at “max” and Muse Spark’s 57 at “xhigh” — each model’s respective ceiling, on ladders of different depth. Second, the ceiling is close but not reached: 56 is one point behind both, and VentureBeat notes Artificial Analysis places Claude Opus 5 at 63 on the same index — the frontier tier remains a different conversation. On AA’s sub-indices, as surfaced through OpenRouter’s AA-sourced panel, 3.7 Flash posts a Coding Index of 76.1 and an Agentic Index of 45.1; Arena.ai’s early human-preference results provisionally rank it ninth overall and eighth for web development. Independent evals exist this time — plural, and broadly consistent with each other.

“Gemini 3.7 Flash costs $0.75 per million tokens” is true on exactly one surface. Google’s official API lists the introductory rate through December 31, 2026, and posts a standard rate of $1.50/$7.50 from January 1, 2027 — the same rate 3.6 Flash’s standard pricing already carries, and 3.6 Flash shares the intro window too, so the cliff as posted today reads Flash-line-wide rather than specific to the new model. The widely repeated “half price” framing is only true against the Flash line’s pre-August-13 workhorse rate. Meanwhile OpenRouter layers its own time-limited 50%-off promotion on top of Google’s already-discounted intro price. Each row below names its surface.

One footnote deserves its own hedge: the model card marks the Flash prices with an asterisk whose footnote text we could not retrieve from the page itself. The introductory-window reading is consistent with Google’s announcement and its official pricing page, so we treat it as the intro-price marker — but it is a presumption, not a verbatim confirmation. And on whether the January 1 doubling actually lands, there is a fresh precedent pointing the other way: Sonnet 5’s own $2/$10 rate was announced at launch as introductory pricing with a scheduled step-up to $3/$15 — and Anthropic cancelled that increase, making the intro rate permanent. Vendors have now demonstrated that a scheduled price cliff is a plan, not a promise. Budget for $1.50/$7.50 from January; don’t be surprised if the cliff moves.

Read as a whole, the nineteen rows don’t crown a single model — they partition the workload space. Here is how we’d translate the full table, the independent AA read, and the surface-labelled prices into routing defaults, treating every margin of two points or less as a tie to be broken by price.

The projection worth making: three-week release cycles mean this table is a snapshot, not a standings board. The durable takeaways are structural — Google is now competitive enough on coding benchmarks to force per-task price comparisons, OpenAI holds the hardest agentic rows, Anthropic holds judgment-heavy knowledge work, and every vendor’s pricing now carries dated conditions that move independently of capability. If your stack still routes every workload to one default model, that is the actual finding of this table — and it’s the kind of routing decision our AI transformation engagements start with, benchmarked on your workloads rather than the vendor’s. For the wider market context beyond these three vendors, see our broader model comparison.