Anthropic, OpenAI, and Google all shipped new “default” model tiers within a six-week window this summer, and by August 2026 the choice between them has real budget consequences. Claude Sonnet 5 went live June 30, GPT-5.6 (in its Luna, Terra, and Sol tiers) hit general availability July 9, and Gemini 3.7 Flash arrived August 13 as Google’s answer to both. Each vendor calls its release the new workhorse tier for coding and agentic tasks, but the pricing gap between them is enormous: Gemini 3.7 Flash costs as little as $0.75 per million input tokens, while GPT-5.6 Sol runs $5.00, a 6.7x spread before you even factor in output tokens.
This comparison breaks down specs, pricing, coding benchmarks, and real deployment patterns for all three models as they stand today, August 25, 2026. If you’re deciding which API to build against, which model to route by default in an agent stack, or whether to switch from an older Claude, GPT, or Gemini tier, the numbers below come directly from vendor model cards, GitHub’s Copilot changelog, AWS’s Bedrock announcement, and independent benchmark trackers.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
Claude Sonnet 5 vs GPT-5.6 vs Gemini 3.7 Flash: Quick Verdict
If you need one sentence: Claude Sonnet 5 wins on raw coding accuracy and is the safest default for agentic engineering work, Gemini 3.7 Flash wins on price-to-performance and is the cheapest way to run high-volume agent loops, and GPT-5.6 (specifically the Terra tier) is the balanced middle option with the deepest tooling ecosystem via OpenAI’s function-calling and computer-use stack. That coding edge shows up on third-party leaderboards too: BenchLM’s aggregate tracker scored Sonnet 5 at 69.8 out of 100, ranking it #18 of 232 models tracked as of September 2026. None of the three is a strict upgrade over the other two — the right pick depends on whether your bottleneck is token cost, coding accuracy, or ecosystem integration.
All three ship within the same release cycle, so this isn’t a case of comparing a flagship against a budget model from a year ago. Claude Sonnet 5, GPT-5.6, and Gemini 3.7 Flash are Anthropic, OpenAI, and Google’s current “everyday” or “workhorse” tier — the model most developers hit by default when they call the API without specifying a premium flagship like Claude Opus 5 or a GPT-5.6 Sol override.
What Changed: The August 2026 Model Landscape
The three releases arrived in a tight sequence that reshuffled the default-tier conversation in under seven weeks:
- June 30, 2026 — Anthropic ships Claude Sonnet 5, immediately setting it as the default model for Free and Pro users on claude.ai, with API access at launch pricing of $2 per million input tokens and $10 per million output tokens, which Anthropic has since made permanent, canceling the planned September 1 increase to $3/$15.
- July 9, 2026 — OpenAI releases the GPT-5.6 family in three tiers: Luna, Terra, and Sol, replacing the previous GPT-5.5 lineup. All three ship the same day into GitHub Copilot.
- August 13, 2026 — Google launches Gemini 3.7 Flash, explicitly positioned by Google’s own blog post as “our most intelligent workhorse model,” aimed at coding and agent workflows, with no announced date yet for a Gemini 3.7 Pro flagship.
Anthropic’s own Claude Sonnet 5 announcement frames the release around agentic coding gains rather than a broad capability jump, which lines up with what Anthropic has done with every Sonnet release since Sonnet 4. OpenAI’s approach is different: rather than one model, GPT-5.6 splits into three price and performance tiers so developers can pick a ceiling that matches their budget, a structure OpenAI first tested with GPT-5.5 and has now carried into 5.6, as detailed in Simon Willison’s write-up of the launch. Google, meanwhile, skipped a Gemini 3.7 Pro entirely for this release and shipped Flash first, according to Google’s own announcement — a sequencing choice that signals Google is prioritizing cost-efficient coding agents over a flagship reasoning jump this cycle.
Context matters here: this is the third major workhorse-tier refresh in about five months across the big three labs, following GPT-5.5, Claude Sonnet 4.6, and Gemini 3.6 Flash. Each new generation has narrowed the capability gap between the labs while widening the price gap, which is exactly the pattern this article’s numbers show.
Full Specifications Comparison Table
| Spec | Claude Sonnet 5 | GPT-5.6 (Terra) | Gemini 3.7 Flash |
|---|---|---|---|
| Developer | Anthropic | OpenAI | Google DeepMind |
| Release date | June 30, 2026 | July 9, 2026 | August 13, 2026 |
| Context window | 1,000,000 tokens | 1,050,000 tokens | 1,000,000 tokens |
| Max output tokens | 128,000 | 128,000 | Up to 128,000 (64K on some routes) |
| Knowledge cutoff | January 2026 (unofficial, not published in model card) | February 16, 2026 | Not publicly specified as of Aug 2026 |
| SWE-bench Verified | 82.1% | Not independently confirmed for Terra tier | Coding benchmark cited at 43.6% by AI Weekly’s internal suite |
| Terminal-Bench 2.1 | 80.4% | Not published for Terra separately | 85.8% |
| OSWorld-2.0 (computer use) | Not published in retrieved sources | Supported via computer-use tool | 47.9% |
| Input modality | Text, image | Text, image | Text, image (coding/agent focus) |
| Output modality | Text, tool calls | Text, tool calls, function calling | Text, tool calls |
| Native audio/video I/O | Not documented | Not documented | Not documented |
| Available via | Claude API, Claude Code, claude.ai, AWS Bedrock, Vertex AI, Microsoft Foundry | OpenAI API, ChatGPT, GitHub Copilot | Gemini API, AI Studio, Vertex AI, Gemini Enterprise |
| GitHub Copilot support | GA since June 30, 2026 | GA since July 9, 2026 (all three tiers) | Not confirmed in Copilot as of Aug 25, 2026 |
The context window gap is essentially a rounding error — all three sit around the 1-million-token mark, which has become the industry standard for this tier of model in 2026. Anthropic’s own documentation confirms Sonnet 5’s 1,000,000-token context window and 128,000-token max output, figures locked into its spec sheet as of January 2026 and still the ones independent trackers cite today. The bigger differentiator is what each lab chose to optimize: Anthropic leaned into agentic coding accuracy, OpenAI split into tiers so buyers can dial reasoning depth up or down, and Google chased low latency and low cost for high-volume coding loops.
Pricing Breakdown: The 6.7x Gap
Pricing is where these three models diverge the most, and it’s not close. OpenAI’s GPT-5.6 family alone spans four price points once you count all three tiers, while Gemini 3.7 Flash undercuts everything on the table during its introductory window.
| Model / Tier | Input ($/1M tokens) | Output ($/1M tokens) | Cached input ($/1M tokens) | Pricing window |
|---|---|---|---|---|
| Claude Sonnet 5 | $2.00 | $10.00 | Not published | Permanent (originally an intro rate through Aug 31, 2026, later made permanent) |
| Claude Sonnet 5 (planned standard, canceled) | $3.00 | $15.00 | Not published | Was to take effect Sep 1, 2026, but Anthropic canceled the increase and made $2/$10 permanent |
| GPT-5.6 Sol | $5.00 | $30.00 | $0.50 | Standard, no promo |
| GPT-5.6 Terra | $2.50 | $15.00 | $0.25 | Standard, no promo |
| GPT-5.6 Luna | $1.00 | $6.00 | $0.10 | Standard, no promo |
| Gemini 3.7 Flash (intro) | $0.75 | $3.75 | Not published | Through Dec 31, 2026 |
| Gemini 3.7 Flash (standard) | $1.50 | $7.50 | Not published | From Jan 1, 2027 |
Run the math on a typical agentic coding workload — say, 500,000 input tokens and 100,000 output tokens per day, a realistic volume for a small team running an AI code-review bot continuously. On Claude Sonnet 5’s introductory pricing, that’s $1.00 plus $1.00, or $2.00 a day. On GPT-5.6 Sol, the same workload costs $2.50 plus $3.00, or $5.50 a day, nearly triple. On Gemini 3.7 Flash’s introductory rate, it drops to $0.375 plus $0.375, or $0.75 a day. Scale that to a month of continuous operation and the gap becomes concrete: roughly $60 for Claude Sonnet 5, $165 for GPT-5.6 Sol, and $22.50 for Gemini 3.7 Flash. That’s the 6.7x-to-7.3x spread that shows up in the headline number for this comparison.
The catch with Gemini 3.7 Flash’s pricing is the expiration date. Google’s introductory rate runs only through December 31, 2026, after which the price doubles to $1.50/$7.50 on January 1, 2027. That’s still cheaper than Claude Sonnet 5’s $2/$10 rate per token, but the roughly 2.7x blended-cost gap between the two narrows once Gemini’s promotional window closes. Sonnet 5’s rate isn’t going anywhere: Anthropic made the $2/$10 pricing permanent on August 10, 2026, canceling the planned September 1 increase to $3/$15 — a move that confirmed the rate Constellation Research’s July 2026 pricing feature had already reported holding through at least August 31, 2026. That makes Gemini 3.7 Flash the side of this comparison still worth budgeting around. Lock in current usage estimates before its promotional rate lapses at the end of 2026, since Sonnet 5’s price is now fixed and GPT-5.6’s tiers carry no published promotional expiration at all.
Coding and Agentic Benchmark Results
Coding accuracy is the benchmark category buyers care about most for this tier of model, since all three vendors are explicitly marketing these releases as developer- and agent-focused. The clearest apples-to-apples number available is SWE-bench Verified, the industry-standard test that measures whether a model can resolve real GitHub issues end-to-end.
Claude Sonnet 5 posts 82.1% on SWE-bench Verified, according to independent benchmark tracker Hokai — a strong result that puts it within range of Anthropic’s own Opus-tier flagship despite Sonnet 5 costing a fraction of Opus 4.8’s per-token rate. On the tougher SWE-bench Pro benchmark, Anthropic’s own July 2026 figures put Sonnet 5 at 63.2%, a number Apidog’s benchmark write-up corroborated, Tech Insider’s July 2026 coverage flagged as 5.1 points above Sonnet 4.6’s score on the same test, and a separate July 2026 MarkTechPost benchmark piece placed just behind Opus 4.8’s 69.2% on that same test. On Terminal-Bench 2.1, which measures multi-step terminal and tool-use tasks, Sonnet 5 scores 80.4%, up from Sonnet 4.6’s 67.0% and ahead of Opus 4.8’s 74.6% per a July 2026 BuildFastWithAI review, a 13.4 point jump in one release cycle — though an earlier June 2026 review from Cosmic JS had clocked the gain at +20.7 points to 76.1% before later tracking settled on the higher figure. Caylent’s July 2026 launch analysis adds a data point Anthropic hadn’t previously published on its own model card: 81.2% on OSWorld-Verified, a computer-use benchmark, which the same MarkTechPost July 2026 write-up noted was an improvement over Sonnet 4.6’s 78.5% on that test. BuildFastWithAI’s July 2026 review also put Sonnet 5 slightly ahead of Opus 4.8 on GDPval-AA v2, 1,618 Elo versus 1,615, even as a separate July 2026 Tech Insider analysis of the Artificial Analysis Intelligence Index scored Sonnet 5 at 57 against Opus 4.8’s 61 — a reminder that Sonnet 5 doesn’t lead every composite benchmark, just the coding-specific ones.
Gemini 3.7 Flash actually edges out Sonnet 5 on the same Terminal-Bench 2.1 test, scoring 85.8% according to Google’s own model documentation. But that number drops sharply on the harder Terminal-Bench 3.0 variant, down to just 14.9%, a gap researchers flagged as a useful reminder that headline scores on an easier benchmark version don’t always hold up on a tougher one. Gemini 3.7 Flash also posts 47.9% on OSWorld-2.0, a computer-use benchmark, and around 72.2% on GDM-MRCR v2 long-context retrieval at 128k context.
GPT-5.6 is the hardest of the three to benchmark cleanly, because OpenAI split it into Luna, Terra, and Sol, and independent SWE-bench numbers specific to each tier were not consistently published across trackers as of this writing. What is confirmed: OpenAI’s own GitHub Copilot changelog describes Sol as having “the highest reasoning ceiling in the family,” implying Terra and Luna trade some accuracy for lower cost and lower latency, consistent with how OpenAI structured the GPT-5.5 tiering before it.
Benchmark Summary Table
| Benchmark | Claude Sonnet 5 | GPT-5.6 (tier) | Gemini 3.7 Flash | Source |
|---|---|---|---|---|
| SWE-bench Verified | 82.1% | Not consistently published | Not directly comparable metric published | Hokai model tracker |
| Terminal-Bench 2.1 | 80.4% | Not published per-tier | 85.8% | Google model docs, Codersera |
| Terminal-Bench 3.0 | Not published | Not published | 14.9% | Datanorth.ai |
| OSWorld-2.0 (computer use) | Not published | Supported (Sol tier) | 47.9% | Datanorth.ai |
| GDM-MRCR v2 @128k | Not published | Not published | 72.2% | Datanorth.ai |
| Prior-gen comparison (Terminal-Bench 2.1) | 67.0% (Sonnet 4.6) | — | — | Codersera launch guide |
The takeaway from three independent sources — Hokai’s model tracker, Google’s own published documentation, and Datanorth.ai’s coverage of the Gemini 3.7 Flash launch — is that Claude Sonnet 5 leads on the benchmark most directly tied to real-world software engineering (SWE-bench Verified), while Gemini 3.7 Flash leads on raw terminal and tool-use speed at the easier benchmark tier, but loses that edge sharply once the benchmark gets harder. Neither GPT-5.6 tier has a fully independent SWE-bench number in wide circulation yet, which is itself a data point: OpenAI has leaned more on internal reasoning-level framing (Sol vs. Terra vs. Luna) than on publishing a single external leaderboard score for the family.
Real-World Adoption: Where Each Model Is Actually Deployed
Benchmark scores matter less than where a model actually ships in production tooling. Here’s what’s confirmed as of August 25, 2026:
- GitHub Copilot added Claude Sonnet 5 the same day it launched (June 30, 2026), available to Copilot Pro, Pro+, Max, Business, and Enterprise users, per GitHub’s official changelog.
- GitHub Copilot also added all three GPT-5.6 tiers (Sol, Terra, Luna) on their July 9, 2026 launch day, per GitHub’s changelog for that release — making GPT-5.6 and Claude Sonnet 5 the two fastest same-day Copilot integrations of any model release this year.
- Amazon Web Services made Claude Sonnet 5 available on Bedrock the same launch day, spanning North America, South America, Europe, and Asia Pacific regions, according to AWS’s own announcement.
- Claude Code, Cursor, and VS Code all shipped Claude Sonnet 5 support at or shortly after launch, per independent coverage from Codersera’s launch guide.
- Google’s Vertex AI and Gemini Enterprise Agent Platform listed Gemini 3.7 Flash as generally available on its August 13, 2026 release date, per Google’s own documentation, positioning it specifically for coding and automated business-task agents rather than consumer chat.
The pattern across all three: this generation of models is being adopted first by coding tools, not consumer chat products. Every confirmed deployment example above is a developer tool — Copilot, Cursor, VS Code, Claude Code, or a cloud agent platform. That tracks with how all three vendors marketed the releases: Anthropic called Sonnet 5 its most agentic model yet, OpenAI split GPT-5.6 into reasoning-level tiers aimed at automation budgets, and Google explicitly framed Gemini 3.7 Flash around coding and agent workflows with no mention of a consumer-facing use case in its own announcement.
Five Real-World Use Case Examples
Here’s how the three models map to concrete engineering scenarios based on their confirmed specs and pricing:
- High-volume CI/CD code review bot. A team running automated pull-request review on every commit across a busy monorepo needs low per-call cost above all else. Gemini 3.7 Flash’s $0.75/$3.75 introductory rate makes it the clear pick, since the bot may process tens of thousands of diffs per month and cost compounds fast at GPT-5.6 Sol’s $5/$30 rate.
- Long-running autonomous coding agent (SWE-bench-style task resolution). For agents that need to independently plan, edit, and verify multi-file changes with minimal human review, Claude Sonnet 5’s 82.1% SWE-bench Verified score and 80.4% Terminal-Bench 2.1 score make it the safer default, even at two to three times the per-token cost of Gemini 3.7 Flash.
- GitHub Copilot enterprise rollout. Since both Claude Sonnet 5 and all three GPT-5.6 tiers are natively available in Copilot today, teams standardizing on Copilot can A/B test Sonnet 5 against GPT-5.6 Terra directly inside the IDE without extra integration work, while Gemini 3.7 Flash isn’t yet confirmed inside Copilot as of this writing.
- Budget-constrained startup MVP. A small team prototyping an AI feature with unpredictable usage patterns benefits from GPT-5.6 Luna’s $1/$6 rate, cheaper than Claude Sonnet 5’s standard tier and simpler to reason about than juggling Gemini’s promotional pricing cliff at the end of 2026.
- Regulated enterprise with existing cloud commitments. Organizations already standardized on AWS Bedrock or Google’s Gemini Enterprise Agent Platform will likely default to whichever model is native to their existing procurement — Claude Sonnet 5 for Bedrock shops, Gemini 3.7 Flash for Vertex AI or Gemini Enterprise shops — rather than making a cross-cloud switch purely for benchmark deltas.
Migration Guide: Switching Between the Three Models
If you’re currently running an older model — Claude Sonnet 4.6, GPT-5.5, or Gemini 3.6 Flash — and want to migrate to one of the three covered here, follow this sequence to avoid breaking production traffic.
Step 1: Audit your current token usage and cost baseline
Pull 30 days of API logs and calculate your actual input-to-output token ratio. Migration decisions change completely depending on whether your workload is input-heavy (large codebases, long context) or output-heavy (long generated responses, verbose agent reasoning traces). Gemini 3.7 Flash’s advantage shrinks on output-heavy workloads since its output price, while still the cheapest of the three, is proportionally closer to the others than its input price is.
Step 2: Update your model identifier string
// Claude Sonnet 5 (Anthropic API / Bedrock)
model: "claude-sonnet-5"
// GPT-5.6 family (OpenAI API) — pick your tier
model: "gpt-5.6-sol" // highest reasoning ceiling
model: "gpt-5.6-terra" // balanced
model: "gpt-5.6-luna" // lowest cost
// Gemini 3.7 Flash (Gemini API / Vertex AI)
model: "gemini-3.7-flash"Step 3: Re-test your system prompts against the new model
Anthropic explicitly deprioritized legacy benchmarks like MMLU, HumanEval, and GSM8K for Sonnet 5’s model card, treating them as saturated in favor of SWE-bench, GPQA, and ARC-AGI — though third-party trackers filled the gap anyway: SkyCrumbs measured Sonnet 5’s July 2026 update at 85-87% HumanEval pass@1, up from roughly 82% at the June launch, even as Anthropic itself stopped reporting the metric. That reporting shift often correlates with subtle behavior changes in how the model handles ambiguous instructions, so run your existing eval suite against Sonnet 5 before flipping production traffic, rather than assuming prompt compatibility with Sonnet 4.6.
Step 4: Recalculate rate limits and concurrency settings
All three vendors cap max output around 128,000 tokens per call (Gemini 3.7 Flash may cap lower at 64,000 on some routes per independent tracker AI Weekly), so agent loops that expect very long single-response outputs need chunking logic regardless of which model you pick. Confirm your retry and backoff logic still works against the new model’s documented rate limits before a full cutover.
Step 5: Roll out with a canary percentage, not a full switch
Route 5-10% of production traffic to the new model for at least a week, comparing output quality, latency, and actual per-request cost against your baseline before committing fully. This is especially important for GPT-5.6, where you’re also choosing between three tiers — a canary rollout lets you compare Terra against Sol on your actual workload rather than guessing from published benchmarks alone.
Pros and Cons: Claude Sonnet 5
Pros: Highest confirmed SWE-bench Verified score of the three (82.1%); largest single-generation Terminal-Bench jump (67.0% to 80.4%); day-one availability across Claude Code, Cursor, VS Code, GitHub Copilot, and AWS Bedrock; introductory pricing undercuts GPT-5.6 Terra on output cost.
Cons: At $2/$10 per million tokens, now Anthropic’s permanent rate rather than a temporary introductory one, Sonnet 5 is still the second-most expensive of the three on both input and output price, behind GPT-5.6 Terra; no published MMLU or general-reasoning benchmark for direct comparison against GPT-5.6 or Gemini; unofficial knowledge cutoff reporting (January 2026) isn’t confirmed in Anthropic’s own model card.
Pros and Cons: GPT-5.6
Pros: Three-tier structure (Luna, Terra, Sol) lets teams pick a cost-to-capability ratio instead of one-size-fits-all pricing; Luna tier is the cheapest input price of any model here at $1.00 per million; largest context window of the three at 1.05 million tokens; native support for function calling, web search, file search, and computer use across all tiers.
Cons: No independently confirmed SWE-bench Verified score per tier as of this writing, making direct coding-accuracy comparison harder; Sol tier is the most expensive output price on this list at $30 per million; three separate SKUs adds decision overhead compared to a single default model.
Pros and Cons: Gemini 3.7 Flash
Pros: Cheapest input and output pricing of the three by a wide margin ($0.75/$3.75 introductory); highest Terminal-Bench 2.1 score of the three (85.8%); strong long-context retrieval performance (72.2% on GDM-MRCR v2 at 128k).
Cons: Score collapses on the harder Terminal-Bench 3.0 variant (down to 14.9%), suggesting the easier 2.1 number may overstate real-world reliability on complex tasks; no Gemini 3.7 Pro flagship announced yet, so there’s no higher-capability tier to fall back to within the same generation; not yet confirmed in GitHub Copilot as of August 25, 2026; introductory pricing expires end of 2026, after which the price doubles.
Latency and Reliability Considerations
Beyond raw benchmark scores, latency matters for anything user-facing. Gemini 3.7 Flash is explicitly built as Google’s speed-optimized tier, following the naming convention Google has used since Gemini 1.5 Flash — Flash models trade some depth of reasoning for lower time-to-first-token and higher throughput. That positioning explains why Gemini 3.7 Flash wins Terminal-Bench 2.1 (a task category that rewards fast iterative tool calls) but drops sharply on Terminal-Bench 3.0, which likely weights harder, more deliberative multi-step reasoning where slower, deeper models like Claude Sonnet 5 tend to hold up better.
GPT-5.6’s tiered structure directly targets this same tradeoff from the other direction: instead of one model tuned for speed, OpenAI lets you choose Luna for latency-sensitive, high-volume calls or Sol for maximum reasoning depth on harder single-shot tasks, with Terra sitting in between. That flexibility is a genuine advantage for teams running mixed workloads, since you can route simple autocomplete-style calls to Luna and route complex multi-file refactors to Sol from the same API without switching vendors.
Enterprise and Compliance Availability
For teams with existing cloud commitments, deployment surface matters as much as raw specs. Claude Sonnet 5 has the widest confirmed enterprise footprint of the three: it’s available on Anthropic’s own API, AWS Bedrock, Google’s Vertex AI, and Microsoft Foundry simultaneously, giving procurement teams flexibility to route through whichever cloud contract already carries their committed spend. GPT-5.6 is available through OpenAI’s own API and, by extension, any Azure OpenAI Service integration that mirrors OpenAI’s model catalog, plus native GitHub Copilot access for all three tiers. Gemini 3.7 Flash is available through Google’s Gemini API, AI Studio, Vertex AI, and the newer Gemini Enterprise Agent Platform, but as of this writing has no confirmed multi-cloud presence outside Google’s own infrastructure the way Claude Sonnet 5 does.
That multi-cloud availability is a meaningful differentiator for regulated industries where data residency or existing vendor agreements dictate which cloud a workload can run on. A healthcare or financial services team already committed to Azure, for instance, has an easier path to GPT-5.6 than to Gemini 3.7 Flash, while a team on AWS Bedrock gets same-day access to Claude Sonnet 5 without any additional procurement cycle.
Multimodal Capabilities: What’s Missing
All three models are marketed primarily around text and code generation, and the multimodal story is thinner than you might expect for a 2026-era release. Claude Sonnet 5 accepts text and image input but outputs text and tool calls only, with no native audio or video input or output documented in Anthropic’s public materials. GPT-5.6 follows the same pattern across all three tiers: text and image input, with function calling, web search, file search, and computer-use tool support layered on top, but no confirmed native audio or video modality. Gemini 3.7 Flash is explicitly scoped around coding and agent workflows rather than broad multimodal input, and Google’s own documentation does not spell out a full audio or video I/O matrix for this release.
That’s a meaningful signal about where this generation of “workhorse” models is headed. Two release cycles ago, model announcements from all three labs leaned heavily on multimodal demos (voice conversations, video understanding, image generation). This cycle, none of the three headline releases lists a major multimodal upgrade as a launch highlight. Instead, all three vendors chose to compete on coding accuracy, tool use, and price. If your use case depends on native audio transcription, video understanding, or image generation baked into the same model call, none of these three releases moves that needle, and you’ll still need a separate specialized model or a multimodal flagship tier like Claude Opus 5 or a dedicated Gemini vision model to cover that ground.
Image input is where all three do agree: each model can read screenshots, diagrams, and photos as part of a prompt, which matters for use cases like debugging a UI from a screenshot, reading a whiteboard sketch of an architecture diagram, or parsing a scanned document. But treat “vision” here as read-only image understanding, not generation. None of Claude Sonnet 5, GPT-5.6, or Gemini 3.7 Flash generates images natively as part of this release, so image-generation workloads still route to separate models entirely (Claude doesn’t generate images at all, GPT-5.6 would route to a separate OpenAI image model, and Gemini would route to Google’s Nano Banana or Imagen family instead of Flash).
How This Generation Compares to Its Predecessors
Judging these three models in isolation misses the more useful signal, which is the rate of change from the generation each one replaced. Claude Sonnet 5’s jump from Sonnet 4.6 is the most dramatic of the three by the numbers available: Terminal-Bench 2.1 climbed from 67.0% to 80.4%, a 13.4 percentage point gain in a single release cycle, which is a larger single-generation jump than Anthropic posted between Sonnet 4 and Sonnet 4.6. Vellum’s June 2026 analysis went further, finding Sonnet 5 ahead of Sonnet 4.6 on 100% of Anthropic’s published benchmarks — a clean sweep rather than a mixed scorecard. That pace of improvement is notable given Anthropic actually cut Sonnet 5’s launch price to $2/$10 per million tokens on June 30, 2026, per Anthropic’s own pricing page, down from Sonnet 4.6’s $3/$15, and locked that lower rate in for good on August 10, 2026, canceling a planned increase back to $3/$15. Anthropic delivered a meaningfully better model while cutting the price, not raising it.
GPT-5.6’s generational story is different: rather than one model improving, OpenAI restructured the entire lineup into three named tiers (Luna, Terra, Sol) replacing whatever equivalent structure existed under GPT-5.5. That’s a strategic shift as much as a capability one — it signals OpenAI sees more value in letting developers choose a price-to-reasoning tradeoff than in shipping a single flagship number to beat. Whether that approach produces better real-world outcomes than a single unified model remains to be independently tested at scale, since granular per-tier benchmark data is still sparse across trackers as of mid-August 2026.
Gemini 3.7 Flash’s generational jump is the most benchmark-visible of the three because Google published a direct comparison in its own model evaluation documentation: the coding benchmark AI Weekly tracked rose to 43.6% from whatever score the prior Gemini 3.6 Flash (released just three weeks earlier, on July 21, 2026) had posted. That three-week gap between 3.6 Flash and 3.7 Flash is unusually tight even by 2026’s accelerated release cadence, and it suggests Google is iterating on the Flash tier faster than either Anthropic or OpenAI is iterating on their respective workhorse tiers, likely because Google has not yet shipped a Gemini 3.7 Pro to absorb some of that development attention.
Which Model Should You Choose?
Based on the confirmed specs, pricing, and benchmark data above, here’s the practical breakdown by team profile:
- Choose Claude Sonnet 5 if coding accuracy on complex, multi-step tasks is your top priority and you can absorb a mid-range cost, especially if you’re already on AWS Bedrock or using Claude Code, Cursor, or GitHub Copilot.
- Choose Gemini 3.7 Flash if you’re running high-volume, latency-sensitive agent loops where per-call cost dominates your budget, and you can tolerate somewhat lower reliability on the hardest tasks.
- Choose GPT-5.6 Sol if you need the deepest reasoning available in this generation and cost is a secondary concern, particularly for one-off complex tasks rather than continuous high-volume calls.
- Choose GPT-5.6 Terra if you want a balanced default that sits between Sonnet 5’s cost and Gemini Flash’s speed, with the added benefit of OpenAI’s mature function-calling and computer-use tooling.
- Choose GPT-5.6 Luna if budget is your primary constraint and your tasks are simple enough that a lighter reasoning tier is sufficient — it’s the cheapest input price on this entire comparison.
The honest answer for most engineering teams in August 2026 is to run all three behind an abstraction layer and route by task type: use Gemini 3.7 Flash for bulk, low-stakes calls, Claude Sonnet 5 for anything touching production code review or autonomous agent work, and GPT-5.6 Terra or Sol for tasks that need the deepest single-shot reasoning. None of the three has emerged as a universal winner, which is itself the headline result of this comparison.
Common Migration Mistakes to Avoid
Teams migrating between these three models tend to make the same handful of mistakes, based on patterns visible across the launch-week feedback captured in independent trackers like Codersera and Hokai. The first is assuming price parity across input and output tokens. Because output tokens cost 5x the input rate on Claude Sonnet 5, roughly 6x on GPT-5.6 across all three tiers, and 5x on Gemini 3.7 Flash, a workload that shifts toward longer, more verbose responses (common when a model starts producing chain-of-thought-style reasoning traces) can blow past a cost estimate that was based on input-token pricing alone. Always model your migration cost estimate using your actual historical output-to-input ratio, not an assumed 1:1 split.
The second common mistake is assuming every introductory pricing window behaves the same way, either always temporary or always permanent. Gemini 3.7 Flash ships with a genuinely time-limited discount, expiring December 31, 2026, after which the price doubles to $1.50/$7.50. Teams that build cost projections off the launch-week rate card without flagging that expiration date routinely get surprised by the jump when the promotional window closes. Claude Sonnet 5 went the other way. Its $2/$10 introductory rate was originally set to expire August 31, 2026 and rise to $3/$15 — a cutoff Constellation Research’s July 2026 pricing feature had also flagged as the rate holding through — but Anthropic canceled that increase on August 10, 2026 and made $2/$10 the permanent rate instead. Build your cost model around each vendor’s currently confirmed rate, and revisit it only when a vendor actually publishes a pricing change rather than assuming one on a guessed timeline.
The third mistake is picking a GPT-5.6 tier based on benchmark scores alone, since independently verified per-tier SWE-bench or Terminal-Bench numbers for Luna, Terra, and Sol are still thin across trackers as of mid-August 2026. Rather than assuming Sol is worth its 2x cost premium over Terra purely because it’s marketed as having “the highest reasoning ceiling,” run your own canary test against your actual task distribution before defaulting to the most expensive tier. In many workloads, Terra’s balance of cost and capability turns out to be sufficient, and the incremental accuracy gain from Sol doesn’t justify the price difference for anything short of the hardest reasoning tasks.
Frequently Asked Questions
Is Claude Sonnet 5 better than GPT-5.6?
On the one directly comparable metric available — SWE-bench Verified — Claude Sonnet 5’s 82.1% score is the strongest confirmed number among the three models, since no independently verified per-tier SWE-bench score has been published for GPT-5.6 as of August 2026. But “better” depends on the task: GPT-5.6 Sol is marketed as having the highest reasoning ceiling in its family, and Terra offers a lower-cost middle tier that Sonnet 5 doesn’t have an equivalent for.
How much does Gemini 3.7 Flash cost compared to GPT-5.6?
Gemini 3.7 Flash’s introductory price is $0.75 per million input tokens and $3.75 per million output tokens, compared to GPT-5.6 Sol’s $5.00/$30.00 — a 6.7x input gap and 8x output gap. Against the cheaper GPT-5.6 Luna tier ($1.00/$6.00), the gap narrows but Gemini 3.7 Flash still costs less on both input and output.
What is the context window for each model?
Claude Sonnet 5 and Gemini 3.7 Flash both support 1,000,000 tokens of context. GPT-5.6, across all three tiers (Luna, Terra, Sol), supports a slightly larger 1,050,000-token context window. Max output is capped at 128,000 tokens for Claude Sonnet 5 and GPT-5.6, while Gemini 3.7 Flash’s output cap varies by route, with some trackers reporting 64,000 tokens on certain endpoints.
Does GitHub Copilot support all three models?
As of August 25, 2026, GitHub Copilot supports Claude Sonnet 5 (added June 30, 2026) and all three GPT-5.6 tiers, Sol, Terra, and Luna (added July 9, 2026). Gemini 3.7 Flash is not confirmed as available inside GitHub Copilot as of this writing, though it is available through Google’s own Vertex AI and Gemini Enterprise Agent Platform.
When does the introductory pricing expire for each model?
Not anymore for Claude Sonnet 5. The $2/$10 per million token introductory rate was originally set to expire August 31, 2026 and rise to $3/$15, but Anthropic canceled that increase on August 10, 2026 and made $2/$10 the permanent rate. Gemini 3.7 Flash’s introductory pricing ($0.75/$3.75) still runs on a clock, through December 31, 2026, before rising to $1.50/$7.50 on January 1, 2027. GPT-5.6’s pricing across Luna, Terra, and Sol has no published promotional expiration and appears to be standard, ongoing pricing.
Which model is best for autonomous coding agents?
Claude Sonnet 5 currently has the strongest published results on the benchmarks most relevant to autonomous coding agents: 82.1% on SWE-bench Verified and 80.4% on Terminal-Bench 2.1, up sharply from the prior-generation Sonnet 4.6’s 67.0%. Gemini 3.7 Flash scores higher on Terminal-Bench 2.1 (85.8%) but drops to just 14.9% on the harder Terminal-Bench 3.0 variant, suggesting its edge may not hold on the most complex multi-step agent tasks.
Can I use these models on AWS or Google Cloud instead of the vendor’s own API?
Yes for Claude Sonnet 5, which launched on AWS Bedrock the same day as its Anthropic API release, spanning North America, South America, Europe, and Asia Pacific regions, and is also available via Google’s Vertex AI and Microsoft Foundry. Gemini 3.7 Flash is available via Google’s own Vertex AI and Gemini Enterprise Agent Platform. GPT-5.6 is accessible through OpenAI’s API directly and through Azure-based integrations that mirror OpenAI’s model catalog.
Is there a more powerful version of any of these models coming?
Google has not announced a release date for a Gemini 3.7 Pro flagship as of August 13, 2026, meaning Gemini 3.7 Flash is currently the newest model in that generation without a higher-tier sibling yet. Anthropic and OpenAI already ship higher-tier options in the same generation: Claude Opus 5 sits above Sonnet 5, and GPT-5.6 Sol is already the top tier within the GPT-5.6 family itself.
