GPT-5.6 vs Gemini 3.5 vs Claude Sonnet 5: a capability-based decision framework

The tier mismatch nobody mentions

Nearly every published comparison in this category lines up three model names and scores them against each other. The problem is that the three names sit at different points in their vendors’ own hierarchies, so the comparison measures positioning as much as capability.

The clearest evidence comes from OpenAI itself. On its GPT-5.6 announcement, the company benchmarks its flagship against Claude Fable 5 — not Claude Sonnet 5 — reporting that Sol “sets a new high of 53.6” on Agents’ Last Exam, “eclipsing Claude Fable 5 (adaptive reasoning) by 13.1 points,” and scoring “80, 2.8 points above Fable 5, while using less than half the output tokens” on coding.

That choice of opponent tells you where the vendor believes the tier line falls. Anthropic’s own documentation describes Claude Fable 5 as its “most capable widely released model,” released 9 June 2026 with a one-million-token context window, while Claude Sonnet 5 is positioned as “the best combination of speed and intelligence.”

Meanwhile Google’s position is frequently misreported. Its 19 May 2026 announcement made 3.5 Flash generally available — “available today to billions of people globally” — while saying of the larger model only: “We’re also hard at work on 3.5 Pro. It’s already being used internally, and we look forward to rolling it out next month.” That next month came and went; press coverage through July and August 2026 continued to report 3.5 Pro as delayed.

So a fair reading of GPT-5.6 vs Gemini 3.5 vs Claude Sonnet 5 has to state plainly which tier each name represents, and note that one of the three has no flagship in the comparison at all.

GPT-5.6, Gemini 3.5 and Claude Sonnet 5 all emphasize coding, tools and agentic work. Their official announcements provide useful facts about product structure and availability, but their benchmark claims are produced by the vendors under different conditions. This comparison therefore focuses on selection criteria, not an invented overall score. Read GPT-5.6 vs Gemini 3.5 vs Claude Sonnet 5 as a fit exercise rather than a contest.

How the products are positioned

OpenAI offers GPT-5.6 as Sol, Terra and Luna, with different capability and cost targets plus an ultra setting for coordinated multi-agent work. Google launched Gemini 3.5 Flash across its consumer, developer and enterprise surfaces and said 3.5 Pro would follow. Anthropic positions Sonnet 5 as a more agentic, adjustable-effort model available in Claude products and its API.

Price per token, and why it misleads

Published rates make the tier structure concrete. Per million tokens, as listed by each vendor:

Read down the output column and the mismatch is obvious: Sol’s output rate is three times Sonnet 5’s. Comparing them head to head without saying so is comparing a flagship against a mid-tier model and calling the price difference a finding.

But per-token rates are also the wrong unit for a purchasing decision, and OpenAI’s own claim shows why. If Sol reaches a given coding score “while using less than half the output tokens,” then a three-times-higher output rate applied to half the tokens produces roughly 1.5 times the cost for that task — not three times. Token efficiency partially cancels rate.

The unit that actually matters is cost per successfully completed task, including human correction. A cheaper model that needs a second attempt and ten minutes of review is more expensive than a dearer model that completes on the first pass. Any evaluation of GPT-5.6 vs Gemini 3.5 vs Claude Sonnet 5 that stops at the price list has measured the input to the decision rather than the decision.

Two further pricing notes worth verifying before budgeting. Anthropic recorded on 10 August 2026 that Sonnet 5’s introductory pricing “is now permanent.” And OpenAI describes Sol, Terra and Luna as “durable capability tiers that can advance on their own cadence” — meaning the tier names persist while what sits behind them changes, so a benchmark tied to a tier name ages differently from one tied to a version number.

Shortlist by operational fit

Then test complete tasks

Run the same prompt set with equivalent tools. Include straightforward, ambiguous and adversarial cases. Score factual accuracy, format compliance, tool success, human correction, latency and total cost. Test more than once and keep versions and dates in the report.

For coding, use a real repository and automated tests. For research, verify citations and quoted material. For customer service, check policy compliance and escalation. For agents, include permission boundaries and recovery from partial failure. Done properly, GPT-5.6 vs Gemini 3.5 vs Claude Sonnet 5 resolves differently for different teams, which is the correct outcome rather than a failure of the comparison.

Sources and methodology

Product descriptions come from official OpenAI, Google and Anthropic release posts. RankBoast has not run a normalized three-model benchmark for this article, so vendor results are not combined into a ranking. Pricing and tier positioning were re-verified in August 2026; vendor rates and availability change frequently, so confirm the current model pages before budgeting. Note in particular that Gemini 3.5 Pro had not shipped at the time of that check.

• OpenAI official source
• Google official source
• Anthropic official source