an early adopter’s dilemma

I’m a solo founder in San Francisco, building a privacy-first health platform (status.health) and relying heavily on AI tools to bootstrap prototypes. Six months ago, Anthropic’s Claude AI felt like a godsend. A friendly 100K-token context assistant that could write code, generate documents, and digest huge files without breaking the bank. I eagerly integrated Claude into my workflow, even paying for Claude Pro and later Claude Max to unlock more power.

Today, I’ve hit a wall. Anthropic’s recent pricing and usage policy changes have effectively priced me out of using Claude. Stricter rate limits, steep token pricing for high-context tasks, and revamped model tiers (“Haiku,” “Sonnet,” and “Opus”) have made the model unaffordable and nearly unusable for an indie builder like me. In this op-ed, I’ll break down the timeline of these changes, from quiet nerfs to public caps, and explain how they’ve crushed the value Claude once offered. Along the way, I’ll share what other developers are saying on forums and contrast Anthropic’s approach with OpenAI, Google, and open-source alternatives. The goal is to shine a light on how a promising AI tool lost its way for the little guys, and to plead (even poetically) for a better path forward.

from haiku to opus: claude’s model tiers and token costs

To understand the squeeze on indie users, we need to talk about Claude’s model tiers. Earlier this year, Anthropic introduced three versions of Claude 3: Haiku, Sonnet, and Opus. These correspond to increasing levels of capability (and cost). Claude Haiku is optimized for speed (a smaller model with shorter context), Claude Sonnet balances performance and efficiency, and Claude Opus is the most powerful model with the highest reasoning ability. In practice, Opus can handle the hardest tasks but burns through tokens up to 5× faster than Sonnet. Each tier has its own pricing: as of mid-2025, Sonnet models cost about $3 per million input tokens and $15 per million output tokens, whereas Opus commands a whopping $15 per million input and $75 per million output. Haiku is far cheaper, Claude 3 Haiku was just $0.25 per million input tokens but it’s also less capable.

Crucially, Anthropic phased out the cheaper models as newer ones arrived. The entire Claude 2 series was retired by July 21, 2025, and even the Claude 3 Sonnet (snapshot 2024-02-29) was deprecated then. Developers who had built cost-efficient workflows around Claude 2 or early Claude 3 suddenly had to migrate to Claude 4 models, whether they could afford them or not. (Anthropic did give notice of deprecations in their docs, but if you blinked you missed it.) The result: no more access to the older, budget-friendly models. A startup that was happily using Claude 2.0 on the cheap in June found that by late July their only “replacement” was Claude Sonnet 4 at enterprise pricing levels.

Anthropic also started charging a premium for high-context usage. Claude’s famous 100K-token context window was a selling point for feeding entire codebases or research corpora. But when they expanded Claude’s context to 1 million tokens in August 2025, it came with strings attached: prompts above 200K tokens are billed double for inputs and 50% more for outputs. This long-context feature is in “public beta” but only for top-tier API customers (Tier 4 and custom plans), meaning indie devs on standard plans can’t even access it yet. In theory I could now process 75,000 lines of code in one go, but practically I’m locked out of that capability unless I somehow become a high-volume enterprise client. Anthropic’s own pricing page acknowledges prompt caching and batching as ways to offset these costs, but implementing such tricks is a burden on developers and only necessary because the raw token prices are so high.

march–may 2025: new models launch, indie access lags

At the start of 2025, Claude was riding high on goodwill from early adopters. Anthropic rolled out Claude 3.5 in stages. First Claude 3.5 Sonnet in June 2024, later Claude 3.5 Haiku and Opus, and boasted about “cost-effective pricing” and 200K token context windows. By early 2025, an even more advanced Claude 4 was in the works. In May 2025, at Anthropic’s first “Code with Claude” developer conference, the Claude 4 model family launched, promising state-of-the-art performance on coding tasks. But something troubling happened in this launch: certain independent partners were excluded.

Windsurf, a popular AI coding assistant startup (ironically being acquired by OpenAI), reported that Anthropic cut off their direct API access to Claude models with less than five days’ notice. “Anthropic decided to cut off nearly all of our first-party capacity to Claude 3.x models,” Windsurf’s CEO wrote, saying they had been willing to pay for full capacity but were still dropped on short notice. Even more galling: when Claude 4 launched, Windsurf wasn’t given access to run it, while other big-name coding tools (Cursor, GitHub Copilot, etc.) were granted direct Claude 4 access from day one. Windsurf had to scramble to route Claude through third-party providers at higher cost, warning users of potential outages in their Claude features. Anthropic, for its part, had just launched its own competing product (Claude Code) in February, raising eyebrows about whether they were sidelining a tool that soon would belong to OpenAI.

For indie developers, the Windsurf incident was a red flag: Anthropic could revoke or limit API access abruptly, even for paying partners. As a founder, I empathized. I also had “Claude inside” my product roadmap and suddenly worried if a quiet policy change might yank that away. Windsurf’s CEO noted, “We are disappointed by this decision and short notice.” I’d use stronger words: it’s hard to trust an AI platform that can change the rules overnight. Unfortunately, June 2025 was only a prelude of worse to come.

july 2025: quiet cuts and “usage limit reached” chaos

In mid-July, heavy Claude users (especially those using the new Claude Code tool for programming assistance) began noticing mysterious new limits. Anthropic hadn’t made a big public announcement yet, but suddenly Claude chats were getting cut off early, often with no warning. I experienced this firsthand: long coding sessions that ran fine in June started hitting a brick wall in July, with Claude refusing to continue. Initially I thought it was a glitch, after all, I was paying $200/month for Claude Max (10-20× capacity) and expected some serious usage headroom. Instead I got cryptic errors and stalled conversations.

I wasn’t alone. On Reddit, users voiced shock and anger at how abruptly Claude became unusable for sustained work. “100% and no warning you are getting close,” one user complained, describing how chats would end abruptly with no chance to even get a summary or hand-off. “It has become impossible to get any actual work done because I am spending all my time explaining what I am trying to do over and over to full chats, and limits hit in a few hours,” they wrote, adding that it had gotten “worse in the last week than anytime in the past 6 months!”. Another frustrated developer vented, “The other day I hit my limit on premium after less than 10 min. The f— am I paying for?”. This sentiment was widely echoed: many of us felt like we were paying for a “premium” service that suddenly wouldn’t let us work for more than a few minutes at a time.

So, what changed? It turns out Anthropic had quietly tightened the rate limits on Claude.ai usage, specifically targeting Claude Code sessions using the Opus model. Without telling users upfront, they introduced hidden caps to stop what they considered abuse (like running Claude 24/7 or sharing accounts). In practice, those hidden changes hit normal users hard. One major change was how Claude chose between Opus and Sonnet models on the Max plan. Previously, a savvy user could stick mostly to the mid-tier Sonnet (to stretch their usage) and only invoke Opus for especially tough tasks. Sometime in July, Anthropic changed Claude Code’s behavior so that Max users’ sessions would start on Opus by default, only switching down to Sonnet after a certain threshold of usage. This “auto model switching” was presumably meant to preserve experience (giving you the best model at first) while preventing you from accidentally running Opus non-stop. But it backfired badly. Users who didn’t manually override the model would burn through their token allotment at 5× the normal rate with Opus, hitting the limit shockingly fast. “I noticed an update made it use Opus first before going to Sonnet. Now when I start Claude I make sure it’s set to Sonnet. Opus is stupid expensive,” one user reported. Others on the $200 Max tier found even that wasn’t enough: “with $200 I’m reaching Opus limits pretty fast, while a week ago that would be possible with parallel instances only,” said one, noting they had to micromanage multiple instances before and now even one instance choked.

By July 17, the situation had blown up enough that TechCrunch ran a piece about it titled “Anthropic tightens usage limits for Claude Code — without telling users.” According to that report, some developers found it “impossible to advance [their] project since the usage limits came into effect.” Anthropic’s response at the time was cagey. They acknowledged being aware of issues but “declined to elaborate further.” In other words, no clear apology or detailed explanation to users. Many of us felt blindsided. We had been early Claude champions, some even canceling ChatGPT Plus subscriptions because Claude’s bigger context and reliable output seemed worth it. Now Claude was throttled and sputtering; it “has been worse in the last week than anytime in the past 6 months,” as one user observed ruefully. The lack of communication and transparency was perhaps the worst of it. “No warning… Wtf,” one user wrote about the sudden capacity errors. “I thought it was just me… I had this insane hallucination that something was changed,” mused another, only half joking. We now know something was definitely changed, just mostly behind closed doors.

august 2025: weekly caps and the $500++/month paywall

After the outcry in July, Anthropic finally went public with a plan: they would impose new weekly usage caps on all paid tiers of Claude.ai, effective August 28, 2025. An email went out to subscribers and Anthropic even tweeted about it, framing the move as targeting a “handful of users” who run Claude Code 24/7 or resell access. They assured us this would affect “less than 5% of subscribers” and that “most users won’t notice a difference”. But the details told a different story for power users and indie devs pushing the limits.

5-Hour Windows + New Weekly Limits: Anthropic kept the existing 5-hour rolling rate limit (the one that limits how much you can do in any 5-hour period) and stacked two weekly limits on top. Now, in any 7-day span, you can hit a hard ceiling where Claude refuses work until the week resets. There are two weekly caps: one on overall usage (hours of any model) and one specifically on Claude Opus 4 usage. This was essentially Anthropic saying: we will now meter out Opus, our “most advanced” model, by the hour. If you used too much Opus time in a week, you’re done until next week (unless you pay more).

What the Caps Mean in Practice: In their announcement, Anthropic gave rough numbers. On the $20/month Pro plan, you can expect 40–80 hours of Claude Sonnet 4 per week (and note: Pro users cannot access Opus at all under the new rules). On Max $100 (5×), they said 15–35 hours of Opus 4 and 140–280 hours of Sonnet 4 per week. On Max $200 (20×), 24–40 hours of Opus 4 and 240–480 hours of Sonnet 4 per week. These might sound like big numbers (who codes 40 hours straight on Opus in a week?), but they aren’t as generous as they look. Those hour figures are not literal clock hours, but based on token usage translated into an “effective hours” metric and Anthropic has been very fuzzy about how they calculate that. Heavy coding tasks with large codebases can chew through an “hour” of Opus much faster than 60 minutes of real time. Users quickly pointed out that the Max 20× plan no longer gives 4× the value of Max 5× once you hit the weekly cap, you pay double, but you might end up with the same weekly Opus allowance if you’re not careful. “20× feels like marketing if a weekly cap cancels it,” as one user succinctly put it.

Extra Usage Now Costs Extra: Anthropic’s message to us was essentially “if you hit the cap, you should go pay for API calls.” They noted Max users could purchase additional usage beyond the weekly limit at standard API rates. In the Claude Code interface, if you reach your allotment, it will helpfully prompt you to switch to pay-as-you-go credits (which cost the same steep per-token prices given earlier). For an indie dev, this feels like moving the goalposts: we signed up for a flat-rate service because we needed predictable costs. Now, if we actually use what we need and exceed some opaque weekly threshold, we’re asked to start swiping the credit card per call. A few of us joked darkly that at the rate things are going, Anthropic would prefer we shell out $500 or $1000 a month just to use Claude at the level we used to on a $100 or $200 plan.

The developer community’s reaction to the weekly caps was scathing. A Reddit megathread filled with nearly a thousand comments got boiled down into a user-generated “discussion report,” and the top issues were lack of transparency and punishing loyal users. “The issue isn’t just limits – it’s opacity,” the report’s TL;DR read, noting that without a live usage meter or clear definitions of what an “hour” means, users get surprise lockouts mid-week. People begged Anthropic for basic quality-of-life fixes: “Give us a meter so I don’t get nuked mid-sprint,” one user quote reads. It’s hard to overstate how disempowering it is to be in the middle of a coding session or a long conversation and have Claude abruptly refuse to continue because you unknowingly crossed some invisible line. Another user pointed out how the 20× Max plan felt like a bait-and-switch: “If 20× doesn’t deliver meaningfully more weekly Opus, rename or reprice it.” Right now, paying for the highest tier doesn’t guarantee you proportionally higher real usage, because the weekly cap “step up” isn’t 1:1. Meanwhile, those of us on the receiving end of these changes feel like we’re being punished for the abuses of a few. Sharing Claude accounts or running bots 24/7 is against Anthropic’s usage policy. Fair enough. But why not enforce those rules on the offenders rather than impose blanket limits on everyone? “Don’t punish everyone – ban account-sharing and 24/7 botting,” as one commenter wrote. Or in another user’s blunt plea: “Ban abusers, don’t rate-limit paying devs.”

Anthropic defends the weekly caps as necessary for reliability, they even cited that Claude Code had suffered multiple outages in the prior month due to overwhelming demand. I don’t doubt that a small percentage of super-users were hammering Claude. The problem is the cure is worse than the disease for those of us who legitimately relied on Claude’s former flexibility. Imagine a freelancer or tiny startup who might have one intense week of work (a deadline sprint where they need Claude’s help refactoring a codebase or iterating on a long document) followed by a light week. Under the new regime, that intense week will almost certainly trigger a lockout, because the weekly cap does not care that you have a quiet week to balance it out. There’s no rollover of unused capacity. No way to request a temporary boost except paying per-token. As one user noted, “Locked out till Monday is brutal. Smooth it daily.” suggesting that a daily limit or some kind of smoothing would be more forgiving than a hard weekly wall. But right now, if you hit your weekly max on a Thursday, that’s it – no Claude for you until the next week.

All of this has effectively erected a $500/month or more paywall for getting the most out of Claude. Why starting with $500? It’s an estimate of what an indie developer would need to spend to regain the freedom they had before. Currently, I have to spend on the order of $800/month. The $200 Max plan will not last you a full week of heavy development now, so you’d need to either maintain multiple subscriptions (risky and against ToS) or start buying API credits on top. For some, even a single $200 plan is already a stretch; $500 or more is out of the question. The conclusion is that Claude is no longer indie-friendly. It’s been groomed for enterprise budgets and tightly metered for everyone else. A founder friend of mine quipped that Anthropic must assume anyone using Claude at scale “has venture funding or a corporate expense account.” The rest of us are left picking up the crumbs or looking elsewhere.

unusable for code, long context, or “artifacts”

The following policy changes have rendered Claude nearly unusable for several of its flagship use cases that drew me to it in the first place.

Long-form Coding Assistance: Claude Code was supposed to be a game-changer for developers introducing an AI pair programmer with an immense context window to hold your entire project. In reality, with the new limits, Claude can’t effectively sustain a coding session on a non-trivial repository. Either you hit the 5-hour window cap (which stops you every so often), or now the weekly cap shuts you down entirely. “I had to increase my plan to continue without interruptions… My fear is that soon I’ll have to sign 50×, 100×… That’s f—ed up, right?” said one user on the Max plan when the limits started biting. Even for those willing to pay more, Claude’s assistance has become stop-and-go. It’s hard to build momentum or dive deep into a coding problem when you’re constantly watching a meter (that you can’t even see) or re-explaining context after a reset. As one Redditor lamented, “it used to be able to follow instructions and now it can’t.” The continuity of thought, arguably Claude’s strength over shorter-context models, is lost.

Artifact Generation: In mid-2024 Anthropic introduced “Artifacts” which is a feature where Claude output things like code files, documents, or website designs into a side pane, letting you interact with them separately. It was a great idea, turning Claude from just a chat into a collaborative workspace. But today, artifact generation is hamstrung by the limits. Users report that artifacts sometimes stop appearing or updating properly due to the new restrictions. “All of a sudden I’m seeing this weird non-updating to the artifacts,” one user noted, suspecting “something was changed.” They were right because behind the scenes, Anthropic likely tweaked how artifact outputs count toward usage, possibly to prevent people from using Claude as a free code generator. The result for me has been incomplete or missing artifact outputs in Claude Code. For instance, I had Claude working on a multi-file Python script; it started writing an simulations.py artifact but hit a limit mid-generation. The artifact never finalized, and I couldn’t retrieve what partial code it had written. When I tried to coax Claude to resume, it acted confused (likely because the context got cut). This never used to happen before. Artifacts were one of Claude’s most promising features for builders, and now they’re another casualty of the crackdown.

Long Documents and Contextual Analysis: A core promise of Claude was that 100K (and now 1M) token context, meaning you could feed an entire book or a trove of customer chats or a huge log file and get comprehensive analysis. But what’s the point of a large context window if the model taps out after a few responses or arbitrarily truncates its output to avoid hitting some token limit? I’ve seen Claude become increasingly conservative with large inputs, often replying “Let’s summarize in parts” or refusing to ingest a long text it would have cheerfully accepted earlier. In some cases, I suspect Claude’s frontend is preemptively limiting input size or chunking behind the scenes (perhaps this is part of Anthropic’s suggested Retrieval-Augmented Generation workaround). Users on Hacker News noted the irony that Anthropic didn’t raise subscription prices, but instead quietly imposed techniques like forced summarization (RAG) to cope with costs. A developer on Reddit observed: “There were never any actual usage limits; they just said e.g. ‘Claude Max 20x users have 20× capacity’… But then [they] go on to say [in fine print] …” implying the limits were always somewhat opaque. Now it’s clear: high-context usage is the big cost center, and Anthropic has effectively kneecapped it for anyone not paying top dollar. I used to load up entire research papers and have Claude cross-reference them. Now, anything beyond a certain length triggers either a refusal or such a brief summary that I might as well not have bothered. It’s devastating for those of us in domains like healthcare, where lengthy guidelines or datasets were finally within AI’s reach thanks to Claude’s context size.

Claude no longer delivers on its core value propositions for individual builders. The model might technically still be capable of great things, but Anthropic’s policies have put those capabilities behind a glass wall. It’s like being given a sports car and then finding the fuel tank has a tiny restrictor plate where you can press the pedal, but you can’t go far or fast anymore.

community backlash: “this feels like betrayal”

The mood among early Claude adopters ranges from disappointment to outright anger. Many feel that Anthropic cultivated our loyalty with a great product, only to yank the rug out with sudden monetization and usage clamps. “I also feel cheated,” one user wrote, asking for alternatives as Claude’s performance degraded and limits tightened. Another said, “I dropped Anthropic just before the pricing structure changes. I’m [switching] between ChatGPT, Gemini and NotebookLM,” referencing OpenAI, Google’s Gemini, and a new NotebookLM product, basically anyone else. On Hacker News, the TechCrunch story on Claude’s usage limits spurred comments noting that AI costs should be going down over time, instead of going up. “Costs should be going down. Going up will reduce adoption by developers hoping to keep their apps low cost,” one commenter noted, pointing out the obvious: as model providers find efficiencies (and as competition heats up), we expect better pricing or higher quotas, not stricter ones. Yet Anthropic’s move was the opposite, perhaps because their own cloud costs or profit goals forced their hand. Either way, it left a bad taste.

What stings the most is the lack of warning and dialogue. There was no clear advance notice to subscribers that, for example, “On July 15 we will adjust our usage policies, expect possibly shorter sessions,” etc. The weekly cap announcement came about a month after the quiet changes. I guess always ask for forgiveness instead of permission and all that. As you might expect, early adopters, many of us indie hackers, felt blindsided. We had designed workflows and even products around Claude’s capabilities. Anthropic’s slogan could well have been “let us handle the heavy lifting, you focus on building.” But by not communicating upcoming limits, they hung developers out to dry. One developer wrote on Reddit: “I recently upgraded to the $100 plan and hit my limits super fast today. It drives me crazy that I can’t see any usage metrics.” That was before the weekly cap rollout; even after, Anthropic still lacks an in-app dashboard showing how close you are to those weekly Opus hours or token quotas. People are literally setting up community-made trackers or manually timing their usage. It’s absurd. A high-tech AI service that makes you use a stopwatch and spreadsheet to avoid being cut off. As one frustrated user pleaded, it’s such a simple ask: “Meter, definitions, alerts – that’s all we’re asking.” The opacity is driving us away as much as the limits themselves.

And let’s not forget: some users also suspect Claude’s quality itself has dipped in recent months. Complaints about the model’s output getting “lazier” or not following instructions as well have popped up. Whether this is objectively true or just an illusion (one jokester blamed it on Claude “being French and it’s August” vacation mode), the perception of a nerf is real. When your chat gets cut off repeatedly, you often have to resend or rephrase queries, which can make the model seem dumber (you’re essentially seeing it fail more). There’s speculation that Anthropic might have tweaked the model to be more conservative or shortened its allowed response lengths to save tokens. We don’t know, but the fact we’re wondering highlights the trust erosion. Early Claude was delightfully eager to help produce full, context-rich answers. Late 2025 Claude feels skittish and hamstrung.

One poignant quote from a Redditor summed up the indie builder’s sense of betrayal: “I was using Claude pretty heavily last year… People were complaining daily about it here and Anthropic insisted they hadn’t changed anything.” In their case it was a deja vu due to a similar degradation happened around summer 2024 and was denied. This time, at least, we have concrete policy updates to point to, but the feeling is the same. We trusted Claude; we even paid when it was in beta because we saw its potential. Now many of us feel cut off at the knees, forced to either cough up much more money or lose a tool that had become integral to our work.

comparing options: openai, google, and open source

It’s worth noting that Anthropic’s rivals have taken very different approaches in this period. OpenAI’s GPT-4, for all its own limitations, has kept a relatively stable deal for individual users. For $20 a month, ChatGPT Plus gives you GPT-4 access with a reasonable (and clearly communicated) cap on requests (currently 50 messages every 3 hours) and no surprise weekly limit. You can’t feed GPT-4 100K tokens of context, its max is 8K or 32K if you have the special version, but in practice GPT-4 often feels more available than Claude now. I can chat with GPT-4 continuously without my session mysteriously “filling up.” And if I need the API, OpenAI’s pricing per token is steep but at least they haven’t slapped new caps on the fly. In fact, OpenAI has been lowering some prices (they cut GPT-3.5 turbo costs significantly in 2023, and offered GPT-4 32K context to developers at a premium but with no usage tiers). That’s not to say OpenAI is perfect. They’ve had their own communication lapses and quality adjustments, but from a pure pricing perspective, an indie developer can budget for OpenAI with more confidence. There’s no $500 surprise paywall lurking.

Google’s Gemini (and related products like Bard or PaLM API) is still emerging, but Google seems to be positioning itself as the developer-friendly alternative. For instance, their recent Gemini “Advanced” model (a rival to GPT-4 and Claude) on Google Cloud is rumored to have flexible pricing and potentially larger context windows without exorbitant cost. Google hasn’t yet put Gemini into a consumer-facing $20/month package as of this writing, Bard is free but limited in other ways, however, they’ve heavily invested in AI through products like NotebookLM (an AI notebook for researchers) and vowed not to sting developers with sudden changes. If anything, Google is more likely to offer higher free usage quotas to entice devs into their ecosystem. It remains to be seen, but many of us are exploring these options. One fellow founder told me he’s “mixing and matching GPT-4, Claude, and Vertex AI (Google) to see which is most cost-effective in the long run.” After Anthropic’s moves, Claude is usually the first to be turned off in that mix when cost becomes an issue.

And then there’s the open-source route. Six months ago, I wouldn’t have seriously considered running a local LLM for my use case. Claude and GPT-4 were just too far ahead in quality. But open-source models have rapidly advanced (think Llama 2, Code Llama, etc.), and importantly, they come with no usage shackles. If you have the computing power, you can fine-tune and run these models entirely under your control. Several indie devs I know have started using local 13B-70B parameter models for coding tasks, using techniques like context splitting and vector databases to approximate a 100K context. It’s not as straightforward as using Claude was, and the model quality can be hit-or-miss, but at least it’s predictable cost: once you’ve set up the server, you’re only paying for electricity or cloud GPU time (which you directly control). There’s a philosophical draw to this as well no AI overlord can throttle me if I’m running the model myself. The downside is you might need a $3,000 GPU or a hefty cloud instance, but over the long run that could be cheaper than burning $500 a month on a service. The open-source community is also working on improving context handling (through smarter retrieval) and some models are surprisingly competent at code when fine-tuned.

In comparing these platforms, I realize it boils down to a simple question: who is building for the indie hacker vs. who is building for the enterprise? OpenAI, despite being a big company, still maintains a very large base of individual enthusiasts and developers thanks to ChatGPT. Google is trying to woo developers onto its cloud with flexible AI offerings. Many open-source contributors are indie hackers themselves scratching their own itch. Anthropic, on the other hand, seems to be pivoting hard to enterprise clients, the kind who will pay for government contracts or integrate Claude into Fortune 500 workflows. (They even offered Claude to “all three branches of the U.S. government for $1” in a trial, which tells you whom they want to impress.) There’s nothing wrong with a business pursuing paying customers, but it feels like Anthropic has forgotten the early adopters that helped prove their technology’s value. Indie builders popularized Claude’s strengths online, wrote blog posts praising its large context, built open-source wrappers, and generally gave Anthropic goodwill and mindshare disproportionate to our pocketbooks. To be sidelined now with restrictive policies and effectively told “you’re not our target user anymore.” That hurts.

hope for a better balance

Anthropic’s mission is about creating beneficial AI, and Claude truly is a remarkable achievement on the technical front. It’s because it was so good that these policy changes sting so much. I want to be clear: I’m not angry that Anthropic needs to control costs or prevent abuse. I understand a startup can’t hemorrhage money providing unlimited AI to a minority of super-users but there’s a right way to handle these challenges. The lack of transparency, abrupt implementation, and one-size-fits-all limits are what turned a reasonable business decision into a fiasco for the community. Anthropic could have engaged developers in dialogue (“here’s what we’re seeing, here are some options we’re considering…”). They could have built tools to help users adapt (like proper usage dashboards or smarter switching that doesn’t default to the expensive model first). They could have tailored solutions that target abusers directly rather than clipping everyone’s wings. Instead we got a summer of confusion and frustration, and a future where using Claude feels like walking on eggshells.

As a founder and an early Claude user, I’m writing this in hopes of change. Anthropic stands at a crossroads: it can either double down on being a high-priced enterprise AI vendor or find a way to keep indie innovators on board. I believe there’s value in the latter. After all, today’s scrappy startup could be tomorrow’s big customer, and the innovations independent developers create often expand what’s possible with AI. Anthropic even interviewed me as a power user. We’re partners in the ecosystem. I want to continue using Claude to build but now I’ve gotta rethink my business plan and my entire development pipeline; I’m priced out.

So, Anthropic, if you’re listening: please remember the little guys who got you here. We know running large models isn’t cheap, but meet us halfway with fair pricing, clarity, and respect for our workflows. At the very least, be upfront and honest when you must impose limits. Surprises belong in AI-generated stories. Please keep them out of service terms.

And to end on a note that perhaps the Anthropic team might appreciate, here’s a short poem:

a plea to anthropic

you gave us  
a brain that thought  
late nights felt shorter  
deployments felt lighter  

builders without vcs  
builders without budgets  
felt like you saw us  

then  
you raised the moat  
without warning  
and sat back  

i’m asking  
please  
leave a sliver-open door  
for the lean ones  
the “not-yet-funded” crowd  

we preferred your model  
we still do  
but if we’re priced out  
we’ll build our own steps  
or climb over

you took away  
a quiet kind of equity  
in months  
don’t let it vanish  
completely