Okay so — you've got the project scaffolded, the chat box is open, and you're about to type the thing you're not sure will work. Cursor blinking. Nothing bad has happened yet. And you stop, because somewhere in the corner of the screen there's a number that goes down when you hit send, and the last time you said "let me just try this," it went down and the thing still didn't ship.
That's not a character flaw. That's a rational response to a real cost. I want to show you the actual number behind that flinch, because the number moved under you this month and almost nobody told you.
The date you didn't know you were watching
If you're wiring your own Anthropic key into an agent — Claude Code, your own script, whatever you've got — you've been running on Claude Sonnet 5 since it launched, at $2 per million input tokens and $10 per million output tokens. That price was never framed as permanent. Anthropic's own pricing page called it introductory, good "through August 31, 2026," with a plain, dated cliff waiting on the other side: $3 and $15 per million tokens starting September 1.
So if you'd looked at your usage on July 29 — a month ago, nothing special about the date except that it's a clean "before" — you were iterating at $2/$10 with a 50% jump already circled on the calendar. That's the kind of thing that makes a careful person ration. Not because $2/$10 was expensive, but because you could see it was about to stop being $2/$10, and you didn't know yet whether your idea was worth building twice — once at the old price, once at the new one.
What actually happened
Earlier this month, Anthropic quietly killed that cliff. I checked the current pricing page directly: Sonnet 5 still lists at $2 input / $10 output per million tokens, and sitting right under the table is a note that says the September 1 increase to $3/$15 "will not occur" — the introductory price is now just the price, permanently.
That's one company. It's not happening in isolation. By multiple outlets' reporting, OpenAI cut two of its GPT-5.6 tiers on July 30 — Luna dropped roughly 80%, Terra by about 20% — three weeks after that model family launched at all. Google is reported to have shipped Gemini 3.7 Flash on August 13 at roughly half of what the prior Flash model cost, through the end of the year. None of these labs announced a press conference about it. They just moved the number, because the other two moved theirs first.
That's the "price war" in the headline, and I don't mean it the way a business outlet means it — three companies defending market share. I mean it the way it actually lands on you: the ceiling you were nervously budgeting toward isn't there anymore, on at least one of the tools you're probably using.
Do the math, not the vibe
Here's the part that actually matters, which is smaller than the anxiety makes it feel. Say a "dumb attempt" — a small feature try, a refactor you're not sure about, a "what if I just tried this" — runs about 20,000 input tokens and 3,000 output tokens. That's roughly the size of the worked coding-session example Anthropic itself publishes on that same pricing page, scaled down to one attempt instead of a full session.
input_tokens = 20_000
output_tokens = 3_000
def cost(input_price, output_price, n=1):
return (input_tokens / 1_000_000 * input_price
+ output_tokens / 1_000_000 * output_price) * n
actual = cost(2, 10, 10) # Sonnet 5, permanent rate
cancelled = cost(3, 15, 10) # the Sept 1 rate that got scrapped
print(f"ten attempts, actual rate: ${actual:.2f}")
print(f"ten attempts, cancelled rate: ${cancelled:.2f}")
ten attempts, actual rate: $0.70 ten attempts, cancelled rate: $1.05
Ten throwaway attempts, all failed, all thrown away, cost you seventy cents at the rate that's now locked in — and thirty-five cents more than that is the entire gap you were bracing for. Not seventy dollars. Not seven. Thirty-five cents, spread across ten wrong guesses. Run the same math on Claude Haiku 4.5, the model built for exactly this kind of fast, dumb, disposable iteration — $1/$5 per million tokens — and ten attempts land at thirty-five cents flat, no cliff to worry about at all.
I'm not telling you this to make you feel silly about the flinch. I'm telling you because the flinch was calibrated to a number that's gone, and it hasn't recalibrated yet.
The two knobs worth knowing about
If you want to push the number down further, there are two real levers on the same pricing page, not tricks — just discounts Anthropic already built in. Prompt caching charges 10% of the standard input price on anything the model has already seen once, which matters if your ten attempts share the same project context each time (they usually do). And the Batch API, for anything that isn't waiting on you in real time, cuts both input and output prices in half — on Sonnet 5 that's $1/$5 instead of $2/$10, the same rate Haiku charges at full price. Neither of these requires you to switch models or learn a new tool. They're a checkbox and a different API call.
The flinch was never really about the money
Andrej Karpathy — not a beginner, not remotely metered by a $20 budget — is reported to have said on the No Priors podcast this year that he feels nervous when he has subscription left over, because unused capacity feels like waste rather than savings. If a guy with enterprise-scale compute access describes the exact same scarcity feeling you get staring at a usage bar, the feeling was never really proportional to the dollars. It's a pattern-match: "spending equals risk," left over from a context where that was true, running unexamined in a context where it mostly isn't anymore.
Ten dumb attempts on the tier most of you reading this are actually using cost less than a coffee. The fear was doing math your calculator never got to run.
If you're not metered in API dollars
Same logic applies if you're burning credits inside Lovable, Bolt, or Cursor instead of a raw key — you're just one layer removed from the number I just showed you, because those tools mark up and repackage the same underlying model pricing. I'm not going to hand you a credit-to-dollar conversion for tools I can't watch move in real time; check your own usage page for what a "generation" or a "fast request" actually costs you today versus what it cost you a month ago. The direction on the big-name models most beginners route to has mostly been down or flat this year, but not every model moved that way — some providers, like DeepSeek, raised prices in the same stretch — which is exactly why it’s worth thirty seconds on your own usage page before your next hover-over-send moment, instead of an assumption in either direction.
Why this matters
The whole premise of "just try it and see" — the thing this site keeps telling you to do — only works if trying costs close to nothing when it fails. For most of this year, on most tiers, that's been quietly getting more true, not less, and the compounding effect of three labs racing each other down is that the failed attempt you're scared of right now is cheaper than it's been at any point since these models existed. Beginners specifically benefit from this in a way seasoned engineers with employer credit cards don't feel as sharply: you're the one actually watching the number, which means you're the one actually rationing based on a fear that's now out of date.
Final Thought
Go open your usage page — the Anthropic console, or whatever dashboard tracks your Cursor or Lovable credits — and look at what a single failed attempt has actually cost you this month. Then go type the thing you've been talking yourself out of. If it takes you ten dumb, thrown-away tries to get there, you now know, in real numbers and not in vibes, roughly what that costs on the tier you're on: pocket change, not a decision you need to agonize over before you hit enter.