Three frontier models in one month is a lot, even by 2026 standards. OpenAI’s GPT-5.6 is the worst-kept secret on the internet, Google’s Gemini 3.5 Pro is crawling out of Vertex preview, and Anthropic dropped Opus 4.8 in May while teasing a “Mythos-class” model that’s apparently too capable to ship without extra locks on the door.
If you pay for one of these — or route API spend across all of them — the question isn’t “which benchmark is highest.” It’s “which one do I send each kind of work to, and is it worth switching.” That’s the part most comparison posts skip. So let’s sort it by job.
First, a reality check on what’s actually real, because the hype cycle has gotten ahead of the release notes.
What’s confirmed, what’s rumored, what’s gated
These three are at very different stages, and pretending otherwise will cost you money.
GPT-5.6 is not announced. As of early June 2026, OpenAI has published no system card, no release note, no model page. What exists: a single canary entry for “gpt-5.6” that flickered through OpenAI Codex’s rollout logs and vanished, internal codenames (ember-alpha, beacon-alpha, iris-alpha — the “alpha” suffix is the tell), and a Polymarket contract pricing a June 30 release at roughly 80–89%. The rumored direction is more agentic execution, better tool use, and improved token efficiency over GPT-5.5. Treat all of that as a bet, not a spec sheet.
Gemini 3.5 Pro is real but staged. Google announced it at I/O on May 19 and it’s still in limited Vertex preview, with general availability expected this month. The headline specs: a 2-million-token context window and a “Deep Think” reasoning mode. Pricing isn’t locked — the base-case estimate has it near $2 input / $12 output per million tokens with a surcharge above 200K context, while a more pessimistic read puts it at the usual ~10x-Flash ratio, around $15 / $60. Big spread. Watch the official pricing page before you build a budget on it.
Claude Mythos 1 is the strange one. Anthropic shipped Opus 4.8 on May 28 and, in the same breath, said models at the next capability level “require stronger cyber safeguards before general release” and would reach all customers “in the coming weeks.” A Mythos Preview is already posting numbers (more on that below), but you and I can’t freely buy it yet. Realistic public window is mid-June through late July.
So right now, exactly one of these three — Opus 4.8, standing in for the Mythos generation — is something you can put on a credit card today. Keep that in mind every time someone shows you a three-way bar chart.
GPT-5.6: the agentic execution bet
OpenAI’s recent arc is consistent, so the rumors are easy to read. GPT-5.5 already leans hard into doing work autonomously — calling tools, holding state across long tasks, recovering from its own errors. It tops Terminal-Bench 2.0 for agentic coding at 82.7%, which is the benchmark that actually matters if your agent lives in a shell and runs commands.
GPT-5.6 looks like a sharpening of that edge, not a reinvention. Expect better tool-calling reliability, fewer wasted tokens on the way to an answer, and tighter integration with OpenAI’s own agent surface. If your workflow is “give the model a goal and a toolbox and let it churn,” this is the lineage built for it.
Where I’d be cautious: GPT-5.5 sits at 58.6% on SWE-bench Pro, the harder coding benchmark drawn from live repos with multi-file diffs. That’s well behind Anthropic. A point upgrade to 5.6 won’t close a ten-point gap. OpenAI’s strength is the agentic loop and the ecosystem around it, not raw code-correctness on gnarly problems. Buy it for the orchestration, not the diffs.
Gemini 3.5 Pro: reasoning plus a 2-million-token room
Google’s pitch is “frontier on reasoning and coding at the same time, with the biggest context window in the business.” The 2M-token window is the genuine differentiator — nobody else lets you drop an entire codebase, a quarter of support tickets, or a stack of contracts into a single prompt and ask questions across all of it. For long-context document work, that’s not a nice-to-have, it changes what’s possible.
Deep Think is the other lever. It’s Google betting on test-time compute — let the model grind longer on hard problems instead of answering fast. The catch worth flagging: Deep Think is reserved for the $250/month Ultra tier, not the $20 Pro plan. So the version of Gemini 3.5 Pro most people will touch through a $20 sub is the fast one, not the slow-and-careful one. If you’re comparing it to a competitor’s deepest reasoning mode, make sure you’re comparing the same gear.
The pricing uncertainty is real and it matters. At ~$2/$12 it’s an aggressive value play against Opus-tier pricing. At ~$15/$60 it’s a premium model you’d reserve for hard work. Those are completely different products from a cost-routing standpoint, and Google hasn’t committed. I’d wait for GA pricing before moving production volume.
For a sense of where the cheaper tier already sits, the Gemini 3.5 Flash vs the field breakdown is a useful baseline — Pro inherits a lot of that family’s behavior.
Claude Mythos 1: the coding frontier behind glass
Here’s where the numbers get loud. Opus 4.8 — the model you can buy — posts 88.6% on SWE-bench Verified and 69.2% on SWE-bench Pro, up nearly five points from Opus 4.7. On SWE-bench Pro it leads GPT-5.5 (58.6%) and Gemini 3.1 Pro (54.2%) by roughly ten points. That’s not a rounding-error lead. On hard, multi-file, real-repo coding, Anthropic is alone at the top.
And Opus 4.8 is the second-best Claude. The Mythos Preview already shows 93.9% on SWE-bench Verified, which is why Anthropic is being cagey about general release. Their stated reason is cyber-safety: a model that good at writing and modifying code is also that good at writing and modifying the wrong code, and they want the guardrails in place first. You can read that as genuine caution or as theater — probably some of both — but the practical effect is the same. The best coding model in the world isn’t on the menu yet.
Opus 4.8 also shipped with the things that actually matter day to day: a fast mode running ~2.5x quicker and 3x cheaper than before, dynamic workflows that let Claude spin up multiple subagents at once, and an effort dial so you decide how hard it thinks per request. Anthropic also claims it’s about four times less likely than Opus 4.7 to let a flaw in its own code slip by unremarked. If you’ve spent an afternoon cleaning up confidently-broken generated code, you know why that line is the one I care about most.
The trade-off: Anthropic is the most restrictive on access and historically the priciest at the top end. You pay for correctness. Whether that math works depends entirely on what the wrong answer costs you.
Head-to-head, by the job you’re doing
Benchmarks are inputs, not decisions. Here’s how I’d actually route work as of June 2026.
Multi-file, production coding
Opus 4.8, and it’s not close on the hard benchmark. SWE-bench Pro is the one that mirrors real engineering — live repos, no leaked ground truth — and a ten-point lead there shows up as fewer broken PRs. When Mythos 1 opens up, it likely extends that lead. GPT-5.6 will be the better agent wrapped around the coding (running the loop, managing the terminal), but for the diffs themselves, bet on Claude.
Deep reasoning and analysis
This is the closest fight. Gemini 3.5 Pro with Deep Think (Ultra tier) and the Mythos generation are both built for slow, careful thinking. If your reasoning task also needs to chew on a huge amount of source material at once, Gemini’s 2M window tips it. If it’s hard but compact, Claude’s correctness edge tips it back. Test both on your actual problems — synthetic reasoning benchmarks lie about this constantly.
Long-context document work
Gemini 3.5 Pro, on the strength of the 2M window alone. Nothing else lets you reason across a whole codebase or a giant document set without chunking and stitching. Chunking is where accuracy goes to die, so removing it is worth real money.
Agentic, tool-calling execution
GPT-5.6 (or 5.5 today). The agentic loop — tool use, state, error recovery — is OpenAI’s home turf, and GPT-5.5 already leads Terminal-Bench 2.0. Claude’s new subagent workflows are a genuine challenge here, so this gap is the one I expect to narrow fastest.
Raw cost per token
Genuinely unknowable until Gemini 3.5 Pro’s GA pricing lands. If it comes in near $2/$12, it’s the value frontier and a lot of volume should move to it. If it’s $15/$60, it joins Opus in the “reserve for hard tasks” bucket and Flash-tier models keep the high-volume work. Don’t architect around a rumor.
Stop betting on one model
The mistake I see teams make is treating this as a single-winner election. It isn’t. The right setup in 2026 is a router: cheap, fast models soak up the volume — classification, drafting, simple extraction, anything you run thousands of times — and you escalate to a frontier model only when a task is genuinely hard or expensive to get wrong.
Concretely, that looks like a Flash-tier or mini model handling 80–90% of calls, with hard coding routed to Opus 4.8 (then Mythos when it opens), long-context analysis to Gemini 3.5 Pro, and agentic execution to the GPT line. You’re not picking a favorite. You’re matching each task to the model that’s best and cheapest for that task, and swapping pieces as prices move.
If you’re wiring this up, an LLM gateway makes the routing layer something you configure instead of hardcode, which matters a lot in a month where three providers are all shipping at once.
So which subscription do you buy?
Depends on who you are.
Solo developer who lives in code: Claude. Opus 4.8 today, Mythos when it lands. The correctness edge on real-repo work pays for itself the first time it doesn’t hand you a plausible-looking bug. The effort dial and fast mode make the cost more controllable than the sticker suggests.
Startup team shipping fast across mixed work: don’t buy one, build the router. Pay for Claude API access for the hard coding, keep an OpenAI key for agentic workflows, and pilot Gemini 3.5 Pro the moment GA pricing is public. The flexibility is worth more than the loyalty discount.
Enterprise already locked into a cloud: lean into the lock-in. If you’re a Google Workspace and Vertex shop, Gemini 3.5 Pro’s integration and context window will outweigh a few benchmark points. If you’re on Microsoft, the OpenAI line plugs in with the least friction. The “objectively best model” matters less than the one your data and tooling already live next to.
The one thing I wouldn’t do is sign an annual commitment this week. Three models are moving at once, GPT-5.6 isn’t even confirmed, Gemini’s pricing is a coin flip, and the best Claude is still behind glass. Run your own tasks through whatever’s available, route accordingly, and give it another month before you bet the budget. Pick one workflow you run constantly and benchmark it across the two or three you can access today — that single test will tell you more than any leaderboard.