English

Based on "LIVE VIBE CHECK: GPT-5.5 Has it all" from Every Watch the original video

OpenAI's Silent Revolution: GPT-5.5 Emerges as the AI Workhorse You Didn't See Coming

A new titan has entered the artificial intelligence arena, and according to a panel of seasoned AI experts, it's not just a contender – it's a game-changer. GPT-5.5, OpenAI's latest model, has been put through its paces in a rigorous three-week "vibe check," revealing a surprising blend of senior engineering prowess and everyday utility that has some calling it their new "daily driver." While not yet available on the API, its rollout to Codex and ChatGPT promises to reshape how developers and knowledge workers interact with AI.

The consensus from the Every.to team, a subscription service dedicated to staying at the forefront of AI, is clear: GPT-5.5 is "really great," "super collaborative," "super fast," and even "pretty personable." But its true distinction lies in a rare duality – its ability to excel at highly complex, senior-level engineering tasks while simultaneously serving as an indispensable workhorse for a broad spectrum of daily activities.

The Workhorse Unveiled: Benchmarking a New Era

Dan, the host and primary tester, unveiled the headline finding: "GPT-5.5 is OpenAI's workhorse model." What makes it so? The team's proprietary senior engineer benchmark, designed to evaluate a model's capacity to rewrite an existing codebase with the insight and skill of an experienced human engineer, delivered astonishing results. Tested against a real codebase, rewritten independently by two senior engineers, GPT-5.5 achieved a best score of 62 out of 100. In stark contrast, Opus 47, a highly regarded competitor, scored a mere 33. This nearly 30-point swing signifies a monumental leap in AI's ability to tackle sophisticated code refactoring.

However, a fascinating caveat emerged: GPT-5.5 achieved its peak performance when guided by a plan literally written by Opus 47. This intricate interplay suggests a compelling division of labor: Opus excels at high-level strategic planning, offering "terse, spec-like, and contract-heavy" blueprints (e.g., "take this gigantic file and get it down to 500 lines"), while GPT-5.5 demonstrates an unparalleled capacity for relentless, assertive execution.

Dan elaborated on this dynamic: "If you give it a prompt that says, 'Hey, I want you to like go and rewrite a major part of this codebase... figure out how you would rewrite it from first principles and then do it,' both 4.7 and 5.5 can figure out... a plan for it." But when it comes to execution, Opus 47 often "gets distracted and a little bit almost intimidated by a really big codebase and a big rewrite," opting for minor patches. GPT-5.5, on the other hand, possesses the "courage or assertiveness" to "delete a bunch of code and... think about this from first principles without getting as distracted by the existing code." This tenacious execution, sustained over "many turns, over many many many hours, over many many tokens," is a novel capability not seen in previous models.

A Model for Every Task? Diverse Perspectives from the Front Lines

The team's diverse roles and workflows offered a multifaceted "vibe check," revealing how GPT-5.5 adapts to different professional needs.

Reliability for the Enterprise: Mike Taylor's Perspective

Mike Taylor, Every's Head of AI Consulting, found GPT-5.5 to be "the most reliable model I tested." Comparing it to a "safe Waymo" versus Opus's "dangerous Tesla," Mike highlighted its dependability for tasks where he couldn't afford to micromanage. For instance, creating curriculum for corporate training materials, which involves sifting through "tons of call notes from different people across the organization" to ensure all themes and issues are represented, became a seamless process.

"It just like hasn't failed on that once," Mike stated, contrasting it with Opus, which, despite producing "really sharp stuff" and "cool titles," often required line-by-line review. For corporate training, "you don't want cool titles all the time... you actually want something dependable, reliable that like, you know, regular people will find accessible." GPT-5.5 proved ideal for tasks needing to be "unoffensive," "reliable," and comprehensive without being "a little bit wild."

The Generalist vs. The Specialist: Kieran Classen's Nuanced View

Kieran Classen, GM of Cora and creator of Compound Engineering, offered a more nuanced take. While acknowledging GPT-5.5's impressive coding capabilities—scoring identically to Opus 4.7 on his specific coding benchmark—he views it as more of a "specialist" than a "generalist."

"Claude is the generalist that is a very good coding model, but it's also very good at product work... looking at the big picture," Kieran explained. GPT-5.5, conversely, "is very good in execution and going into details but sometimes it like breaks down if you look at it from far away and you just see things not being coherent." As a "product engineer generalist" building a new version of Cora, which involves wide-ranging work from front-end to back-end, he still reaches for Opus 4.7 as his daily driver.

Kieran also noted a significant limitation: its performance with Ruby. While "very, very good at... React, like a Next app" and TypeScript, he found it "just not good at Ruby," which is Cora's primary language. "That's not how you write Ruby," he lamented, making it a "biggest blocker" for his specific workflow.

The Vibe Coder's Dream: Naveen's "Well-Rounded" Experience

Naveen, GM of Monologue, represents the "engineer engineer" perspective, aligning more with OpenAI's approach. He describes GPT-5.5 as a "really well-rounded model," a significant shift from his previous preference for Claude/Opus for coding. He now uses 5.5 for everything from Python and Swift codebases to support replies for Monologue.

Naveen's most compelling example is his "vibe coding" adventures. While recovering from pink eye, he leveraged 5.5 to create three different apps from scratch, including "Dayline," a Raycast alternative for daily to-dos. He simply gave it a screenshot of Raycast and a high-level idea, and 5.5 "vibe coded" the entire Mac app, including complex minor interactions, in a single thread spanning "200 million tokens." He even expanded it to an iOS app that syncs automatically.

"I didn't look at single code," Naveen emphasized. This incredible feat highlights 5.5's ability to maintain context over extremely long conversations and manage multiple codebases simultaneously, a testament to its robust "compaction" or context management. Mike Taylor echoed this, noting that his own "Karpathy-style knowledge base" project, initially needing a "Ralph Wiggum loop" (a constant human-like check-and-commit process), now runs faster with 5.5 simply compacting itself, requiring "less harness essentially."

Speed as a Superpower: Austin's Shift in Knowledge Work

Austin, Every's Head of Growth, presented perhaps the most dramatic shift in allegiance. Previously a dedicated Claude Code user, he now finds the Codex app, powered by GPT-5.5, to be his "daily driver for everything I do." For someone without a technical background, 5.5 has made engineering work, from building dashboards to shipping landing pages and creating strategic plans, "way more comfortable."

A few months prior, Austin found older Codex models alienating, making him "feel like I'm dumb." The models would ask clarifying questions in a tone that felt dismissive. "All of that is gone for me in the new model," he stated. "I both understand what it's saying [and] I also really trust what it's saying."

The speed of the Codex app with 5.5 is a crucial factor. Austin described pointing the model at Notion and Slack, brain-dumping high-level campaign goals, and receiving a campaign plan that was "90% of what it came up with." This efficiency allows him to manage complex tasks "while I'm in meetings" or "watching the NBA playoffs," nudging the AI along with only "10% of my brain working." This "speed is in some ways... a type of intelligence," Dan added, granting users "a lot more power."

Creative Crossroads: Design and Imagery

While GPT-5.5 excels in many areas, the panel noted mixed results in creative tasks, particularly design. Kieran observed that while typography and structural alignment in 5.5's designs looked better, the overall aesthetic could be "a little bit chaotic," sometimes showing "some degradation" compared to previous models. He found Opus 4.7 still performed better on subjective "cozy island" design tests. Austin also found Opus models superior for video clipping and image generation via tools like ThreeMotion, noting 5.5's "laziest possible approach" of simply zooming in on recordings.

However, a new contender has emerged from OpenAI itself: the new GPT image generation tool. Both Dan and Austin were "blown away" by its capabilities. Austin, initially hesitant, used it to create the YouTube thumbnail for their very live stream in "like 3 minutes," requiring "no notes." It produced high-quality, clean icons, accurate finger counts (without a source image), and even a "95% approximation of our logo." Dan added that it's "very good at like, 'Oh, I want you to change this one little thing, but keep the rest the same,'" and impressively, it can generate accurate likenesses of people. This suggests OpenAI is rapidly closing the gap in creative visual domains.

The Evolving AI Landscape: A Luxury of Choice

The "vibe check" on GPT-5.5 paints a picture of an AI landscape rich with specialized and increasingly powerful models. While GPT-5.5 distinguishes itself as an assertive, high-performance workhorse, particularly in complex engineering execution and fast-paced knowledge work, other models like Opus 4.7 retain their edge in strategic planning and certain creative design tasks.

The experts' varied experiences underscore a crucial point: "All the models are very good... There's nothing bad about any of these models. They're amazing." The choice now often boils down to subtle nuances, specific workflow alignment, and even the language being used. OpenAI's GPT-5.5 has undoubtedly raised the bar, offering a glimpse into a future where AI not only assists but actively drives complex projects with unprecedented speed and reliability, empowering users to achieve more with less friction. The revolution, it seems, is well underway, and it's happening at warp speed.