February 2027: What Actually Changed in AI Tools

The holidays are over. CES is a fading memory. The teams are back at full capacity, and February is the first month where we see what 2027 is actually going to look like — not what the roadmaps promised, but what developers sitting at desks decided to ship. It's also Q1 enterprise buying season, which means a wave of releases optimized to close deals rather than solve problems. Let's sort it out.

What Shipped in February

February was dense. More meaningful releases than any single month since September's conference season, which either means the year is starting fast or everyone had the same idea about timing.

Anthropic shipped Claude Code 2.0 — or more precisely, shipped enough improvements to Claude Code in a single release that the version number doesn't matter as much as the feature list [VERIFY]. The big addition is a web-based interface that runs alongside the terminal version, which means the "I don't live in the terminal" crowd can finally use it without learning to love the command line. The underlying agent loop also got tighter — fewer abandoned attempts, better recovery from errors, and significantly improved handling of large codebases where the previous version would lose context and start hallucinating file paths. If you tried Claude Code six months ago and bounced off it, February's version is a different enough experience to warrant a second look.

OpenAI released what they're calling "Operator 2.0" — their browser-automation agent — into general availability [VERIFY]. Version 1.0 was a proof of concept that demonstrated the idea while failing at enough tasks to be frustrating. The improvements are real: it handles multi-step web workflows (fill form, navigate, extract data, fill another form) with maybe 70% reliability on common patterns. That sounds low, and it is, but it's high enough to be useful for repetitive tasks where a 30% failure rate means you check the output rather than doing the whole thing manually. The gap between "useful despite being imperfect" and "reliable enough to trust unsupervised" is exactly where most AI tools live right now, and Operator is a clean example of the species.

Google released Gemini 2.0 Pro, and this one deserves attention. The jump from 1.5 Pro to 2.0 Pro is substantial across the board — coding, reasoning, long-context analysis, multimodal understanding [VERIFY]. More importantly, the pricing is aggressive. Google appears to have decided that the way to win the model war is to offer comparable quality at lower prices, and if that's the strategy, it's working. Gemini 2.0 Pro sits at roughly 60% of the cost of equivalent Claude or GPT models for most tasks, with performance that's within spitting distance [VERIFY]. The "Google can't do AI" narrative from 2024 is laughable from the vantage point of February 2027. They can, they are, and they're doing it cheaply.

Meta released Llama 4 Scout — the smaller, faster model in the Llama 4 family — and it's genuinely impressive [VERIFY]. Running on consumer hardware with quantization, it outperforms last year's best open-source models at a fraction of the compute cost. The larger Llama 4 Maverick is still in limited release, but Scout alone shifts the calculus for anyone building products on open-weights models. You can now run a model locally that's competitive with GPT-4 Turbo on most tasks. On a laptop. That was science fiction eighteen months ago.

In the image generation world, Midjourney shipped V7 [VERIFY], and the headlines are about the new "style consistency" feature that lets you lock a visual style across generations. The practical impact is that brand-consistent image generation — same lighting, same color palette, same aesthetic treatment across a campaign's worth of images — is now possible without prompt engineering gymnastics. This is the feature that moves Midjourney from "creative exploration tool" to "production tool," and the design teams I've talked to are already restructuring workflows around it [VERIFY].

Enterprise Releases: Who's Targeting Q1 Budgets

Enterprise AI is its own genre. The features are wrapped in compliance language, the pricing requires a sales call, and the marketing materials use the word "governance" more than a political science textbook. But some of what shipped in February matters to people who don't have "procurement" in their job title.

Microsoft pushed a major update to Copilot for Microsoft 365 that — finally — makes the Copilot experience in Office apps feel integrated rather than bolted on [VERIFY]. The improvements are most noticeable in Excel, where Copilot can now handle multi-step data analysis workflows that previously required you to describe each step separately. In PowerPoint, it can generate slide decks from documents and data that are merely bad rather than the previous standard of aggressively terrible. Small victories.

Salesforce shipped Einstein Copilot updates targeting sales workflow automation [VERIFY]. The CRM-aware AI assistant that can draft emails based on deal context, summarize account histories, and suggest next actions. The sales teams I've talked to say it saves genuine time on administrative tasks. The sales teams Salesforce quotes in press releases say it's "transformational." The truth is between those two points, closer to the first.

AWS launched Amazon Q Developer updates with improved code transformation capabilities [VERIFY] — specifically, automated migration from older frameworks to newer ones (Java 8 to 17, .NET Framework to .NET 8, that sort of thing). Migration tooling is boring. Migration tooling that works is genuinely valuable. Early reports suggest it handles straightforward migrations well and complex ones poorly, which is an accurate description of every automated migration tool ever built, with or without AI.

Tools That Promised January But Slipped

The slip list is instructive because it tells you who's struggling, not who's lying. Delays are normal. Serial delays are a signal.

Rabbit R2 was supposed to ship in January. It didn't. The new date is "Q2" [VERIFY], which in hardware-startup speak could mean anything from April to "please stop asking." If it ships and it works, we'll cover it. The V1's credibility deficit makes skepticism the appropriate default.

OpenAI's "full" GPT-5 release — not the Turbo variant but the full model — was expected in January based on their own timeline hints. It's now targeted for "early 2027" which could still mean February or March but is clearly later than planned [VERIFY]. The Turbo version is good. The full version is presumably better. How much better matters because the pricing will be higher.

Several AI agent startups that promised Q4 2026 or Q1 2027 launches for their products have gone quiet. Not dead-quiet, but the kind of quiet where the Twitter account is still active but the shipping updates have stopped. Names you'd recognize if you follow AI Twitter, but that most people have never heard of, which is perhaps the answer to why the shipping updates stopped.

Early-Year Leapfrogs

The competitive landscape reshuffled faster in February than in any month since last September.

Google leapfrogged on price-performance. If you're building a product and your primary concern is cost per output token at acceptable quality, Gemini 2.0 Pro is now the rational default choice [VERIFY]. Claude and GPT remain better on specific tasks — Claude for code and long-document analysis, GPT for broad general knowledge — but the gap isn't wide enough to justify 40-60% higher costs for many applications. Google's strategy of competing on economics rather than trying to win on quality alone is more sustainable than most commentators give it credit for.

Llama 4 Scout leapfrogged the entire "local inference is a compromise" narrative. Running a genuinely good model on your own hardware used to mean accepting significant quality degradation. Scout on recent hardware is good enough that the quality conversation changes from "how much worse is local" to "how much better do I need than local." For privacy-sensitive use cases, air-gapped environments, and latency-critical applications, the question is now answered: you can run it locally and the results are fine.

AI Marketing Claims That Don't Hold Up

February's crop of dubious claims centered on the enterprise launches.

"Copilot saves 30% of work time for knowledge workers." Microsoft published a study with this number [VERIFY]. The methodology involved self-reported time savings from early adopters who opted into the program and had incentives to report positive results. Selection bias doesn't begin to cover it. The honest version: Copilot saves meaningful time on specific, well-defined tasks for people who learn to use it well. It does not save 30% of total work time for the average knowledge worker. If it did, companies using it would have already reduced headcount or output would have measurably spiked. Neither has happened at the scale the number implies.

"Our model achieves human-level performance on [benchmark]." Three different model providers used some version of this claim in February [VERIFY]. The benchmarks in question measure narrow, well-defined tasks. "Human-level" on a math benchmark means "scores the same as an average human on these specific math problems." It does not mean the model can do math the way a human can. The conflation is deliberate and the audience it misleads — executives making purchasing decisions — is exactly the audience the marketing is aimed at.

"AI agents can now autonomously complete complex business workflows." Multiple enterprise vendors shipped this claim alongside their February releases. In testing, "complex business workflow" means "three to five well-defined steps with clear success criteria." That's useful, genuinely. But it's not what "complex" means in any enterprise context where the word matters. The gap between what the marketing says and what the product does is smaller than it was a year ago. It's still large enough to be a problem.

One February Update That Quietly Improved a Daily Workflow

Obsidian shipped an AI plugin that handles local embedding and retrieval over your vault using locally-run models [VERIFY]. No cloud API calls, no data leaving your machine, no subscription beyond Obsidian itself. You ask a question about your notes, and it returns relevant passages with source links.

This matters not because it's technically revolutionary — local RAG over personal documents has been possible since 2024 — but because it's packaged well enough that non-technical knowledge workers can use it. The Obsidian community is large, the use case is obvious, and the implementation is good enough that it works on the first try for most vaults. One of those tools that makes you wonder why it took this long, which is usually a sign that the packaging problem was harder than the technology problem.

Q1 Momentum Check

Is 2027 starting fast or slow? Fast. Unambiguously fast.

The model layer improved across every major provider in the first two months. The tool layer is shipping meaningful features, not just incremental updates. The open-source ecosystem is producing models good enough to build products on. The enterprise market is spending, which means the revenue is there to fund the next round of improvements. The competitive dynamics are healthy — enough players to drive innovation, enough consolidation to drive quality.

The risks are the same ones they've been: the hype-to-delivery ratio remains unfavorable for anyone making purchasing decisions based on announcements rather than shipped products. The AI tool graveyard will continue accepting new residents. The models will continue to hallucinate at rates that make full autonomy unreliable for critical tasks. These are structural constraints, not temporary inconveniences, and anyone telling you they'll be solved by Q3 is selling something.

But the trajectory is real. The tools are materially better than they were in January. The January tools were materially better than the December tools. The compounding improvement is visible month over month now, and it's happening across the entire stack — models, tools, integrations, workflows. February 2027 is the strongest month for actual shipping we've recorded since this series started. Whether that pace holds is the question that matters for the rest of the year.

The Bottom Line

February delivered. The first full shipping month of 2027 produced meaningful improvements across models (Gemini 2.0 Pro, Llama 4 Scout), tools (Claude Code 2.0, Cursor background agents getting better), and enterprise products (Copilot, Q Developer). The competitive landscape is shifting toward price-performance competition, which benefits everyone who actually uses these tools. The marketing claims remain ahead of the reality, but the reality is catching up faster than in any previous year. If this pace holds through Q1, 2027 is going to be the year that the tools become genuinely good enough that the conversation shifts from "should I use AI tools" to "how do I use them well." We might be there already.

This is part of CustomClanker's Monthly Drops — what actually changed in AI tools this month.