OpenAI released GPT-5.6 on July 9, and I immediately did what I apparently do whenever someone gives me a more capable AI model: I started another website project.
This is now my third major personal website build involving AI, alongside a Node-based writing app I’ve been developing for my science fiction project. I build custom WordPress sites, migrate and restructure content, create calculators and features, troubleshoot styling, and keep adding increasingly elaborate systems to what began as a fairly basic chat app connected to an OpenAI API key. GPT-5.6 Sol has been excellent at all of it.
The problem with writing a clean model review is that Claude Fable 5 and Opus 4.8 are also excellent at this kind of work. I use Claude Code primarily for professional WordPress projects. The newly rebranded ChatGPT desktop app, formerly the standalone Codex app, is my go-to for my personal sites and chat app. At the level of complexity I regularly give them, all the newest models can produce very good results.
Someone maintaining a massive repository or delegating extremely long-running autonomous work may see important differences that I do not. OpenAI’s launch materials emphasize Sol’s gains in agentic coding, professional knowledge work, cybersecurity, science, browsing, and computer use, including stronger results than GPT-5.5 on several relevant benchmarks.
For my work, the distinction is less dramatic. That point does not require another thousand words because I recently wrote about the increasingly narrow experiential gap between frontier models. They are all very good now. Moving on.
Creatively, GPT-5.6 is still the model I prefer. I have continued using it for character development, scene exploration, and plotting within my extremely overbuilt science fiction project, and it fits me better than Claude does. I am not sure I see a radical difference from GPT-5.5, although 5.6 gave me an extraordinarily flattering summary of me and all my projects in our first thread, so perhaps the real breakthrough is more advanced personalized ass-kissing.
The response went beyond basic flattery because the model understood how the projects connect. My websites, writing app, AI experimentation, work initiatives, and science fiction planning are not separate hobbies in my head. They are one sprawling system that makes perfect sense to me and probably looks like stimulant-induced evidence-board behavior to everyone else. GPT-5.6 understood the system. I appreciated that. But obviously it recognizes patterns even better than I do.
The more interesting story is not whether Sol gives me 6% more usable WordPress code or understands one additional layer of my fictional character’s trauma. It is how favorable the response to GPT-5.6 has been compared with the growing frustration surrounding Anthropic, and what that contrast exposes about the two companies.
GPT-5.6 Builds on OpenAI’s Recent Recovery
Recent OpenAI model releases have frequently arrived with an immediate wave of complaints.
Users have objected to personality changes, disappearing models, tighter guardrails, colder responses, excessive reasoning, degraded creative writing, and the loss of whatever strange emotional alchemy made an older model feel special. I have participated in some of this complaining myself, though without describing the model as my abusive ex because I retain basic psychological distinctions.
GPT-5.6 has had a smoother reception, but it did not come out of nowhere.
GPT-5.5 had already marked a turning point. It received a strong reception from users who cared about personality, creative flexibility, role-playing, and emotionally continuous conversations. After several releases that felt like capability gains came with greater stiffness or control, 5.5 felt more like a clear win.
That matched my own trajectory. GPT-5.3 and 5.4 worked perfectly well, but they did not inspire much enthusiasm. During that period, Claude Opus 4.6 was becoming my primary writing tool because it could hold an enormous amount of context and engage seriously with the fictional world I had already built.
GPT-5.5 pulled me back toward ChatGPT. GPT-5.6 has continued that momentum rather than resetting the product’s personality and making everyone begin another round of grief counseling for software.
There are certainly users who prefer Fable or think Sol falls short on particular coding tasks. Online model evaluations remain an unstable mixture of benchmarks, screenshots without context, emotional attachment, genuinely sophisticated testing, and people announcing that a model has been lobotomized because it wrote one mediocre paragraph.
Still, the overall response to GPT-5.6 has been notably positive. Sol is powerful, fast enough to use comfortably, available across paid ChatGPT plans, and part of a broader family that includes the lower-cost Terra and Luna tiers. OpenAI also introduced higher reasoning settings and an Ultra mode that can coordinate parallel subagents for more demanding work.
At the same time, OpenAI folded the former Codex app into the new ChatGPT desktop app, combining Chat, Work, and Codex rather than treating coding as a completely separate product.
That change reflects how people already use these tools. Codex was never only useful for traditional developers. OpenAI itself has been expanding it into reports, spreadsheets, presentations, research, automation, and lightweight internal tools.
I use Codex for WordPress sites and my fiction-planning app. Someone else may use it to analyze investments or prepare client deliverables. Calling the whole product ChatGPT makes more sense, even if longtime Codex users were understandably startled when their coding agent changed icons and looked like the chat app ate it.
More importantly, OpenAI is offering Sol as part of its continuing product lineup.
That sounds like a low bar until you compare it with Anthropic.
Fable’s Access Problems Are Making Sol Look Better
Anthropic released Fable 5 as its strongest general-use model, attracted enormous demand, and then announced that included access through ordinary paid Claude plans would end after a temporary promotional period. Continued use would require additional metered spending.
Anthropic initially planned to end the promotion on July 7. It extended the deadline to July 12 and then again through July 19. Then, on July 18, they announced that after July 20, Fable would be available only on Max and Premium Team plans (with users only able to spend 50% of their usage on it) while Pro and Team Standard users would receive a one-time $100 usage credit and would otherwise have to pay for credits instead of having Fable included in their monthly plans. The extensions gave users more time with Fable, but they also made the rollout feel improvised.
Users were being invited to build workflows around a flagship model while repeatedly checking whether it would remain included next week. Anthropic suggested that broader subscription access could return if capacity allowed, but that did not resolve the immediate uncertainty. Some users may have been happy with the July 18 announcement, but others were still frustrated that Fable would not be included at all monthly plan levels.
Model quality matters. So does knowing whether you can continue using the model without opening a second billing account in your soul.
I may believe Fable is slightly better at a particular type of long-horizon planning. That difference becomes less compelling when Sol is already excellent and included in the product I pay for, especially now that I have downgraded to Claude Pro at $20/month from a $130/month Claude Max plan and switched to a $100/month OpenAI Pro plan.
That was not a theoretical judgment about benchmarks. It was a response to access uncertainty, pricing, and the amount of friction involved in using the strongest model.
OpenAI has certainly engaged in opaque routing, model replacement, access restrictions, and maddening product decisions of its own. But OpenAI launched a strong model with continuing access at the exact moment Anthropic was demonstrating how difficult it could be to reliably use the strongest model it had built.
Sam Altman hasn’t held back from pointing that out.
Sam Altman’s Anthropic Jab Worked for a Reason
On July 9, the same day GPT-5.6 became broadly available, Anthropic launched a campaign called “Inviting hard questions.”
The campaign asks who should decide the rules for AI, whether AI can improve children’s futures, whether it makes the world more dangerous, and whether it can help scientists cure diseases. Anthropic says it will gather public input, show its work, and report on what it is doing to address those questions.
On July 14, Sam Altman responded:
“hard questions are great but only if we deem you worthy enough to not silently downgrade you, or even get access at all”
He also said he initially thought Anthropic’s account was satire.
The insult was opportunistic. OpenAI does not have clean hands when it comes to quietly changing model behavior or deciding which users deserve access to which capabilities. Anyone who has tried to determine whether ChatGPT is actually using the model listed in the interface should not accept transparency lectures from either side.
But the jab worked because it targeted Anthropic’s chosen identity.
Anthropic does not market itself merely as a company that builds capable AI systems with strong safeguards. It presents itself as the company willing to ask whether the AI race itself is endangering humanity. Its branding emphasizes moral seriousness, public benefit, constitutional principles, institutional restraint, and a willingness to contemplate limiting powerful systems.
When that company cannot clearly tell paying subscribers whether they will still have access to its flagship model next week, the contradiction becomes easy to mock.
OpenAI can be accused of behaving like an aggressive technology company and mostly shrug. Anthropic wants to be seen as the conscience of the AI industry.
Consciences are held to a more annoying standard.
Claude’s Paternalism Extends Beyond Dark Fiction
I usually discuss Claude’s cautious personality in relation to creative writing because that is where it becomes most obvious for me.
I write dark science fiction. My characters experience mental illness, addiction, violence, coercion, imprisonment, political extremism, and generally terrible decision-making. ChatGPT has historically been better at understanding that I want an energetic creative partner rather than an intervention from a digital social worker.
But I notice the same difference when asking for ordinary management advice.
I lead a team at a technology company, and part of my actual job is driving AI adoption. That would remain part of my job even if I were not personally obsessed with AI and spending my free time building websites at an unreasonable pace.
When I discuss resistance from developers or explain that I demonstrated how AI could perform work they assumed required a much longer process, Claude tends to focus on how intimidating my energy may feel because the tools excite me. It sympathizes heavily with employees who may feel threatened or embarrassed by seeing a manager reproduce parts of their work, even though I am only showing them how they should be working in 2026 rather than attempting to take on their roles myself.
Those feelings may be real. I am not unaware that rapid technological change can be uncomfortable. I also have a job to do.
Showing developers what is possible with AI is not an act of aggression. If I can use Codex or Claude Code to produce a WordPress implementation quickly, the organizational lesson is that developers should become highly capable AI-assisted developers. I should not conceal what the technology can do because competence might be emotionally activating.
I have made it clear that my goal is to build a stronger team whose members use AI, not eliminate the team. Claude still often responds as though my enthusiasm itself is the management problem.
ChatGPT may flatter me more, and I am aware of this risk and have chosen to survive it. But it also tends to align better with how I actually work. I frequently use ChatGPT to develop an idea first, pushing it forward without hesitation, before bringing it into Claude to refine, structure, or synthesize into a formal document. Claude remains very good at that second phase, particularly for design thinking and organizing complicated material, but it is no longer where I start.
At work, I still heavily promote Claude Code because it is genuinely excellent for development. Personally, Claude has narrowed into a more specialized role, while ChatGPT has become my main environment for exploration, iteration, and momentum.
My partner has joked that Claude defends my developers because Anthropic feels guilty about creating the tools that may replace them. That is obviously not a literal explanation. Claude does not possess corporate guilt or employee solidarity.
Still, the pattern is relevant. Anthropic places unusual emphasis on harm prevention, labor disruption, emotional consequences, and the moral burden of deploying powerful systems. It is reasonable to infer that the company’s priorities affect the behavior it rewards in Claude, although a personal interaction cannot prove a direct causal line from corporate philosophy to one management response.
The result can feel thoughtful in one situation and frustratingly paternalistic in another. I do not need an AI model to assume my worst intentions because I am enthusiastic and responsible for making sure my team adapts to reality. Sometimes the people resisting change are simply wrong.
Anthropic Wants to Build the Future and Warn Us About It
Dario Amodei worked on concrete problems in AI safety before Anthropic existed. The company has invested in alignment, interpretability, scalable oversight, model evaluations, and research into how increasingly capable systems could behave in dangerous or unpredictable ways.
Anthropic is also a public benefit corporation with a Long-Term Benefit Trust intended to help its governance account for the broader effects of advanced AI. Former Federal Reserve chair Ben Bernanke recently joined the trust, reinforcing Anthropic’s attempt to position itself as more than a conventional technology company.
It would be lazy to conclude that Anthropic talks about danger only because safety is useful branding.
The company appears to sincerely believe that advanced AI presents severe risks. It also sincerely wants to build the most capable AI systems in the world.
That creates a real tension. Anthropic is racing toward the frontier while arguing that the frontier may require stronger restrictions, government oversight, access controls, and institutions capable of intervening in development.
Its position seems to be that AI progress is extraordinarily valuable and potentially catastrophic, so the right organizations must remain near the frontier to steer it responsibly.
Naturally, Anthropic believes it is one of the right organizations.
This is not necessarily hypocrisy. In some ways, sincerity creates the harder problem. A company that genuinely believes humanity’s future is at stake may also genuinely believe it has a moral obligation to decide which capabilities people can access and what restrictions other developers should face.
That is where effective altruism can no longer be treated as a minor historical footnote.
Effective Altruism and the Case for Controlling AI Progress
Effective altruism began with a reasonable premise: use evidence and careful reasoning to determine how limited resources can do the most good.
The movement initially became known for causes such as global poverty, public health, and animal welfare. A substantial part of it later turned toward longtermism and existential risk. Longtermist reasoning gives moral weight to the enormous number of people who could exist in the future and therefore treats threats to humanity’s long-term survival as unusually important.
Advanced AI became a major concern because even a relatively small probability of extinction or permanent loss of control can dominate an expected-value calculation when the potential harm includes humanity’s entire future.
This logic may not be completely stupid. Low-probability catastrophic risks should not be ignored merely because they have not happened yet. Serious researchers have outlined plausible concerns involving autonomous weapons, cybersecurity, biological misuse, mass surveillance, concentrated power, and systems that behave in ways their developers cannot reliably predict.
My concern is what happens when speculative future catastrophe overwhelms every measurable cost in the present.
Once extinction is placed on one side of the equation, almost anything on the other side begins to look minor. Slowing research, restricting deployment, limiting model access, expanding surveillance, imposing licensing requirements, and concentrating capability inside a few approved institutions can all be framed as proportionate safeguards.
The benefits lost through delayed progress are harder to count. We cannot identify the exact diseases that might have been treated sooner, businesses that would have survived, discoveries that would have happened, disabled people who would have gained more independence, or years of human labor that could have been saved. Those costs appear as futures that never occurred.
A hypothetical extinction event is vivid. Opportunity cost is invisible.
Critics of longtermism have therefore argued that its focus on abstract future populations can distract from present needs and give a small group of institutions disproportionate authority to define what protecting humanity requires. AlgorithmWatch has criticized the assumptions embedded in longtermist AI narratives, while other critics have questioned whether fixation on future superintelligence can hinder progress against more immediate problems.
Anthropic did not invent effective altruism, and not every Anthropic decision should be attributed directly to it. The company did, however, emerge from the broader intellectual and funding ecosystem in which longtermism and catastrophic AI risk became major priorities. Its governance structure and public mission explicitly emphasize the long-term effects of advanced AI.
These are not hidden connections but rather part of the company’s stated identity.
The question is whether that worldview encourages Anthropic to treat broad access and rapid adoption as dangers to be managed by institutions that consider themselves unusually qualified to do so.
AI Safety Can Also Protect the Companies Already Ahead
Some AI capabilities may eventually require serious regulation. Governments already restrict technologies where private misuse could impose enormous public costs, and advanced AI is not automatically exempt because the product comes in a friendly chat window. But regulation also creates barriers to entry.
Large frontier labs can absorb compliance costs more easily than startups, independent researchers, open-source developers, and smaller competitors. A licensing system designed to prevent catastrophic AI can become a moat around the companies already wealthy enough to build it.
The companies warning governments that frontier models are dangerous are also the companies most likely to survive strict frontier-model rules.
Anthropic may genuinely believe powerful AI should be developed only by organizations with rigorous safety systems, interpretability research, deployment controls, and public-interest governance. Conveniently, Anthropic has designed itself to qualify as one of those organizations.
Its technical excellence makes this tension more frustrating. Claude Code is excellent. Fable may be the strongest available American model for some long-running work (Buzz has started forming around Chinese startup Moonshot AI’s Kimi K3 model). Anthropic’s products demonstrate why rapid AI adoption is valuable even as the company warns that capability may need to be restricted by the responsible.
Anthropic clearly does not want to stop AI. Its release schedule, enterprise expansion, infrastructure spending, and pursuit of frontier research make that obvious. It wants to advance AI aggressively while retaining significant authority over how the resulting capabilities are distributed and governed.
Perhaps someone does need to exercise that authority. I remain unconvinced that the correct someone is automatically the company selling the model.
OpenAI Is Not Necessarily the Ideological Champion of Unrestricted Progress, and Yet…
OpenAI has retired popular models, altered behavior without satisfying explanations, used opaque routing, and centralized its strongest capabilities behind expensive subscriptions. It has also advocated for regulation while occupying one of the strongest positions in the market.
At the same time, it is worth acknowledging that OpenAI has recently moved in the opposite direction on one specific axis: guardrails. GPT-5.5 and GPT-5.6 were widely perceived by users as loosening some of the stricter behavioral constraints that had frustrated people in earlier releases. Creative writing became less constrained, role-playing more flexible, and responses less likely to default to refusal or moral framing in borderline cases. OpenAI has not publicly framed this as “lowering guardrails,” but the shift is visible in both user feedback and the models’ behavior.
Anthropic, by contrast, has leaned more explicitly into structured safety. Claude models are built around a “constitutional AI” approach, where the system is guided by a set of principles designed to reduce harmful outputs and encourage cautious reasoning. In practice, this often results in more frequent refusals, more hedging, and a stronger tendency to reframe or redirect prompts that touch on sensitive topics. Many users, including me, experience this as a more heavily guardrailed system, even when the underlying capability is extremely high.
That difference matters for how the products feel. OpenAI’s recent models are more willing to engage directly and push forward with user intent, while Claude is more likely to pause, contextualize, or question it.
But this is not a clean philosophical divide. OpenAI has not abandoned safety, and Anthropic has not abandoned capability. Both companies are constantly adjusting where they draw the line between usefulness and restriction, and both have incentives to move that line depending on user demand, regulatory pressure, and competitive positioning.
Sam Altman mocking Anthropic for gatekeeping is funny because the criticism is well targeted, not because OpenAI has discovered a principled opposition to gatekeeping.
Both companies are trying to build valuable products, influence governments, win enterprise customers, manage dangerous capabilities, and control the direction of a global technology race. Their differences matter, but neither operates outside ordinary incentives involving money, reputation, and power. What differs is the product philosophy users currently experience.
OpenAI is presenting GPT-5.6 as a tool for ambitious work, with fewer visible constraints than some of its earlier models. Anthropic is presenting Fable as an extraordinary capability shaped by a more explicit safety framework, while asking the public who should control AI.
I know which approach is more appealing to me.
Why GPT-5.6 Is Winning This Moment
GPT-5.6 did not need to defeat Fable on every benchmark to win the release cycle.
It needed to be excellent, broadly useful, creatively compatible with existing ChatGPT users, and continuously available through the subscriptions people already purchased. It achieved that while Anthropic repeatedly extended access to its own flagship and launched a campaign about moral responsibility.
The contrast wrote itself. My own subscription change reflects it. I downgraded Claude and upgraded ChatGPT because Sol gives me strong coding performance, a better creative partner, a more useful general environment, and less uncertainty about whether the flagship model will remain available under my plan.
That does not mean OpenAI is permanently committed to access, immune to paternalism, or incapable of retiring a model I love with minimal warning. I have been through enough releases to avoid developing corporate amnesia after one good week.
Anthropic may still have the better model for certain tasks. It may also be sincere about some of the risks accompanying advanced AI. Sincerity does not resolve the power question.
Anthropic’s campaign asks who should decide the rules for AI. That is a genuinely hard question, and the answer cannot simply be whichever frontier company believes most strongly in its own moral judgment.
For now, GPT-5.6 is winning because OpenAI shipped a model people want and broadly lets them use it. That should not feel radical, but this is the AI industry, so apparently it does.
