Four models dropped this week. Your default just got expensive.

Four big model releases hit in one week. Grok 4.5, GPT-5.6, OpenAI's new voice models, and Anthropic's Claude Science all landed on top of each other. If your team is still routing everything to the same model you picked six months ago, you are probably paying too much for work a cheaper model does better. Let me catch you up on what changed and how to check it before your next review.

The price fight is real, and it favors you

The cost gap between these models is wide enough to matter to your bill. SpaceXAI priced Grok 4.5 at $2 per million input tokens and $6 per million output. Anthropic's Opus 4.7 runs $5 in and $25 out. That is not a rounding error. Elon Musk called Grok 4.5 an "Opus-class model, but faster, more token-efficient and lower cost."

OpenAI is playing the same game. It claims GPT-5.6 Sol matches Anthropic's Claude Mythos 5 using a third of the output tokens. Output tokens are what you pay for on the way out, and they add up fast on long jobs.

So the vendors are handing you leverage. A model that does the same work for a third of the tokens changes the math on any workflow you run at volume. But cheaper only wins if the quality holds for your actual tasks, and that is the part nobody can tell you.

Vibe checks lie, so build a bench

Here is the trap. A vendor benchmark tells you a model is good at their tasks, not yours. And a one-off test where you paste in a prompt and eyeball the answer feels useful but you can't compare it to last month.

Claire built a repeatable bench to fix exactly this. She used Claude Code, frozen inputs, and a fixed rubric, then blind-tested Sonnet 5 against Sonnet 4.6, Opus 4.8, GPT-5.5, and Gemini 3 Pro across PRDs, prototypes, and agent tasks. Same tasks every time a new model drops. That is the whole point: results you can line up over time.

The build is not heavy. An HTML scoring page took her about 45 minutes with Claude Code. Claude Code can even read your old session history and suggest eval tasks based on the work your team actually does. That context is sitting on your machine unused.

The judge you trust is a person

The surprise in Claire's bench is what happened when she let models grade the outputs. She had GPT-5.5 and Opus 4.8 judge, and both were too generous. Their scores clustered in the middle. They missed broken prototypes and ignored wireframe rules that she caught in the first screenshot.

Her taste and the automated scores disagreed almost completely. The model judges ranked Gemini 3 Pro highest and Sonnet 4.6 lowest. Her ranking was nearly the reverse. When she scored 64 generations by hand with a simple 1-to-5 gut rating, that human signal turned out to be the most useful part of the whole thing.

What this means for you: do not outsource the final call to a model. Put a person with taste in the loop, and make sure your rubric encodes what your team actually cares about before you trust any automated number.

Winning is per-task, not per-model

Stop shopping for one model to rule your stack. Claire's task-by-task read: GPT-5.5 for PRDs, Sonnet 4.6 for prototypes and chitchat, Opus 4.8 or Sonnet 5 for dense codebase work. Sonnet 4.6 stayed her daily driver for agent work, not because it scored best, but because she liked how it talked to her.

Routing matters more than picking. OpenAI's new GPT-Live-1 voice models send hard queries to GPT-5.5 for reasoning while keeping the conversation going. That split, a fast model up front and a heavy one behind it, is a pattern worth copying in your own product. More than 150 million people already talk to ChatGPT through voice, so the interface bet is not small.

Anthropic went the other way and built a product around one job. Claude Science targets research the way Claude Code targets engineering. Different models are aiming at different work. Match the tool to the task.

The deep cut

The cheap-token pitch is designed to move your default. Grok 4.5 at $2 in and GPT-5.6 at a third of the tokens both read as "switch to us and save." But Claire's bench showed Sonnet 5 finished near the bottom of her preference ranking, which means the cost argument only holds if the quality argument also holds for your use case. You cannot know that from a launch post.

So the concrete move for Monday: spend one afternoon building a small blind bench on the five tasks your team actually ships. Freeze the inputs, score them by hand, and keep it. Then every launch week becomes a quick rerun instead of a guess or a full migration. The teams that have a bench will switch models when it pays and ignore the hype when it doesn't. The teams without one will keep paying incumbent prices out of habit.

Three questions for your team

1. What are the five tasks we run most, and do we have a frozen, repeatable bench for them, or are we still trusting vibe checks?
2. Are we tracking token cost per task? If not, we have no way to prove a switch to Grok 4.5 or GPT-5.6 actually saves us money.
3. Are we routing simple work to a cheap model and hard work to a strong one, or paying Opus prices for jobs a mid-tier model handles fine?