Three Frontier Labs, Three Different Winners. Route Accordingly.

The Bengaluru SaaS founder pulled up the April invoice on Wednesday morning. ChatGPT Plus, two seats, sixty dollars. Anthropic API through Claude Code, four hundred and twelve dollars. Google AI Studio on Gemini 3 Pro, two hundred and eighty-seven dollars. OpenRouter, mostly free tier, eleven dollars. The Plus subscription that used to feel like the centre of the AI stack was now line four, and the engineering lead had already opened the Slack thread asking whether to cancel it. Nobody on the team had used the ChatGPT UI for production work in three weeks.

This is the shape of the operator stack in May 2026. Three frontier labs race on different lanes, and the team that picks one model and lets it ride pays a tax that compounds every sprint.

The Frontier Tier, Verified

OpenAI shipped GPT-5.5 with model ID gpt-5.5-2026-04-23, a one-million-token context window, and an API rate card of five dollars per million input tokens and thirty dollars per million output. Above two hundred and seventy-two thousand input tokens, the session bills at two times input and one and a half times output, so the headline price hides a real surcharge on the long-context path. Cached input drops to fifty cents per million, a clean ninety percent discount that you only collect if you actually architect for cache reuse.

Anthropic shipped Claude Opus 4.7 with the full one-million-token context at standard pricing, five dollars in and twenty-five dollars out per million. No long-context surcharge, prompt caching at ninety percent off, batch API at fifty percent off. The new tokenizer is the gotcha. Same English text consumes up to thirty-five percent more tokens than the previous Opus generation, so your effective bill on identical prompts is higher than the rate card alone suggests. Read your token counts on the new model before you assume the math holds.

Google shipped Gemini 3 Pro at two dollars input and twelve dollars output per million, with a one-million-token context on the standard variant and a two-million-token window on the Gemini 3.1 Pro release that landed in February. That price is less than half of Opus 4.7 on input and output combined, and Gemini's free tier still exists with reduced quota, which matters for experimentation and Indian shops that need to ship a prototype before the budget conversation opens.

Three flagships, three different lanes. The Artificial Analysis intelligence index puts GPT-5.5 at the top by a thin margin on the latest run, with Opus 4.7 within a point and Gemini 3.1 Pro a step behind. Lane-specific benchmarks tell a different story. GPT-5.5 leads Terminal-Bench 2.0 at 82.7 percent for agentic terminal work. Opus 4.7 leads SWE-bench Pro at 64.3 percent for complex coding. Gemini 3.1 Pro leads GPQA Diamond at 94.3 percent for scientific reasoning. Hallucination rate is where the verdict tightens. GPT-5.5 at the highest reasoning setting hallucinates on 86 percent of unanswerable factual prompts. Opus 4.7 at maximum effort sits at 36 percent. Gemini 3.1 Pro lands at 50 percent. The model with the highest leaderboard score is also the model most likely to confidently invent a fact, and a serious operator stack treats that as a routing input.

The Mid-Tier Value Play

Below the flagships, the value tier is where most production traffic should land in May 2026. Claude Haiku 4.5 at one dollar input and five dollars output per million handles classification, extraction, short-form generation, and customer-support routing for the overwhelming majority of traffic that fits under 200K tokens. Claude Sonnet 4.6 at three dollars input and fifteen dollars output is the right call when Haiku's accuracy floor leaks but Opus is overkill on the unit economics. Gemini 3 Flash at fifty cents input and three dollars output is the cheapest credible reasoner in the frontier-adjacent tier, and it is still the only one of the three with a real free quota you can put in front of users. GPT-5.5-mini has not shipped at this rate card on the OpenAI public pricing page, which is itself a signal: OpenAI is keeping the headline model premium and letting GPT-5 Mini at twenty-five cents and two dollars carry the cheap-tier load.

Per-Task Router

This is the section to bookmark. Verified rates as of May 2026, paired with the lane each model actually wins.

The reasoning behind the table is operator math, not marketing copy. Long-context retrieval is the one place Opus 4.7 has a structural moat. GPT-5.5's two-times input surcharge above 272K tokens means a 900K-token RAG call costs roughly twice what the same call costs on Opus, and Opus hallucinates less on the recall path. For an Indian fintech ingesting RBI circulars or a legal-tech shop indexing High Court judgments, that is the routing decision that matters.

Tool use and agentic terminal work is GPT-5.5's lane. The Responses API is more mature than Anthropic's tool-use loop on edge cases, function-calling schemas are more reliable in production, and the Terminal-Bench score is not a vanity metric. For an agency running scraper pipelines or a CI bot that drives shell commands, GPT-5.5 is the default.

Hindi and Hinglish is the lane where Indian operators historically defaulted to GPT-4o on muscle memory. That muscle memory is now wrong. Gemini 3 Pro's native multilingual training and lower output rate make it the cost-per-correct-output winner on code-switched copy, and OpenAI's own IndQA benchmark confirms Hinglish is a real gap on the GPT family. Sarvam-105B sits behind Gemini as the sovereign fallback if data residency forces the question.

The Indian Shop Cost Math

Picture a Delhi content agency running fifteen million Hinglish output tokens per month at a three-to-one output-to-input ratio. The routing decision shows up in concrete rupees. Eleven and a quarter million output tokens on GPT-5.5 at thirty dollars per million is three hundred and thirty-eight dollars, roughly twenty-eight thousand rupees. The same workload on Gemini 3 Pro at twelve dollars per million is one hundred and thirty-five dollars, eleven thousand two hundred rupees. The agency saves sixteen thousand eight hundred rupees per month on output alone, and the Hinglish quality is better. The gap between a fifty-five percent gross margin and a sixty-eight percent gross margin on the same content retainer comes down to one routing call.

A coding agent run over a 400K-token monorepo on Opus 4.7 burns five dollars on input and twenty-five dollars per megatoken out. The same agent on GPT-5.5 above the 272K cliff burns ten in and forty-five out. At a hundred-PR-per-month cadence Opus is the right unit economics, and the SWE-bench delta means fewer review rounds per PR.

The Free Tail Nobody Talks About Loudly

Nous Portal is hosting Qwen 3.6 Plus free during its preview window, a 1M-context model with vision and tool use that scores 78.8 percent on SWE-bench Verified. OpenRouter is hosting DeepSeek V4, Owl Alpha, and Qianfan-OCR on the free tier at 200 requests per day per model. For a startup doing first-pass classification, scrape summarisation, or low-stakes drafting, this is where bulk traffic should land before touching the paid frontier. An operator stacking Cerebras free inference, Nous for long-context free, and OpenRouter for multi-model fallback runs a zero-dollar floor that handles roughly sixty percent of inference volume.

Fallback routing is the operator's discipline. Free tier first. Mid-tier paid on quality miss. Flagship only when the cost-per-correct-output math says yes. Single-model loyalty is a hobby. Routers ship.

Wire Up This Week

First, get API keys for all three frontier labs and at least two free-tier aggregators. OpenAI, Anthropic, Google AI Studio, OpenRouter, Nous Portal. Total time, ninety minutes. Vault the keys before you do anything else.

Second, build a routing layer with three pinned models and a free fallback. Opus 4.7 for long-context and coding, Gemini 3 Pro for vision and Hindi, GPT-5.5 for tool use. Qwen 3.6 Plus on Nous as the zero-dollar floor. The router is forty lines of Python or a config file in your existing LLM gateway. Do not over-engineer it.

Third, instrument the routing decision with token counts and cost-per-task. Log every call with the chosen model, input tokens, output tokens, latency, and a quality flag. Two weeks of data tells you where your defaults are wrong.

Fourth, retire the seat-based subscriptions that are not pulling weight. If your team is not using the ChatGPT Plus UI in production, the sixty dollars a month is a tax. Move it to API spend on the model that actually wins the task. The receipt should match the routing, not the marketing.

The frontier is no longer a single mountain. It is three peaks, and the operator who maps every job to the right peak runs leaner than the one paying for a single all-access pass.

Related

• Codex bridge, GPT-5.5 and Opus 4.7 in one terminal loop
• How I built a 5-model LLM fallback router
• Memoize LLM calls and save API quota with deja