A developer posted his GitHub bill from June 2026. He expected the usual $29. But it was $750. A kind of similar thing happened with another user. All this happened because GitHub switched its billing model.
That is what getting worse looks like in AI software right now. It directly shows up in the invoice, not a press release.
I went through 2026 changelogs, pricing pages, and release notes for a dozen-plus AI tools across productivity, marketing, and business software, cross-checking each claim against a second source before writing it down. Some AI tools quietly shipped real upgrades. A few quietly made themselves worse, usually through pricing, not features. Here’s what actually changed, with the receipts.
Key Takeaways
- Perplexity’s Comet browser went free on March 18, 2026, after being paywalled.
- Notion cut AI agent running costs by 35-50% while shipping Custom Agents.
- Otter.ai cut its Pro transcription allowance from 6,000 minutes to 1,200 minutes a month at the same $16.99 price.
- GitHub Copilot moved to usage-based billing on June 1, 2026; some bills rose more than 20x.
- Salesforce has cycled through 4 Agentforce pricing models since late 2024, and adoption has stayed low as a result.
How I Tracked What Changed in AI Tools This Year
I didn’t run a lab test on 40 AI tools. All I had to do was pull the vendor changelogs and pricing pages directly, then check each against independent trackers.
I also weighed this against 3 rankings readers may already know: DataNorth’s top 10 for 2026, ranked by raw capability, Dan Martell’s business-growth tier list, ranked by founder ROI, and Bumppy’s roundup of tools creators use.
4 AI Productivity Tools That Quietly Got Way Better in 2026
Longer context windows and cheaper inference are the boring explanation for why productivity tooling moved the fastest in 2026. The interesting part’s where the savings actually landed.
Perplexity: Comet went free, and the Max tier got real
Perplexity made its Comet browser free from March 18, 2026. They removed the paywall that had kept it behind their Pro version. Pro cost around $20 a month, but they launched a new Max tier priced at $200 a month with its own credit system for heavier research work.
DataNorth describes Perplexity as routing subtasks across more than 20 frontier models inside auditable sandboxes. Dan Martell’s list ranks it C tier for limited RoI. So, in the end, the model rewards research-heavy work and underwhelms for quick answers.
Cursor: The router is doing the actual work
Cursor still runs a credit-pool model: Pro at $20 a month with a $20 credit pool, Pro+ at $60, Ultra at $200 with a $400 pool. What changed is the Cursor Router, which picks the right model per request instead of defaulting everything to the most expensive one.
DataNorth measured this as cutting costs by 30-50% versus a single-model approach. A router that stops sending simple autocomplete requests to your priciest model is the difference between a $20 bill and a $60 one.
Claude: The adoption numbers are the real story
The adoption numbers stand out whenever I look at Claude. Its SWE-bench Verified score is more than 80%, and the developer adoption has reached 43%. But Anthropic’s approach feels different. It combines an effort ladder with task budgets, letting me cap how much autonomous work an agent uses before checking in.
For me, that helps explain Claude’s momentum among AI productivity tools, and how the Copilot story above became possible.
Notion: Cheaper agents, gated access
Notion is the most mixed entry here, because the improvement and the catch are the same change. In April 2026, Notion cut AI agent running costs by 35 to 50%. On May 4, 2026, it shipped Custom Agents, measured at $10 per 1,000 credits.
The catch: Notion retired its standalone AI add-on in 2025, so full AI access now requires the Business plan at $20 per member per month. Agents got cheaper to run; getting into the room where you can use them got pricier. Dan Martell’s list puts Notion AI in F tier, suggesting this fits teams already on Notion, not a standalone AI purchase.
3 AI Marketing Tools That Made Real Gains in 2026
Native brand controls and better generation quality did more for AI marketing tools this year than any single flashy launch.
Canva: Magic Studio picked up five real features in one month
Canva’s May 2026 release bundled five upgrades rather than one headline feature: Magic Eraser now removes shadows and reflections along with the object itself, Image to Video animates human faces realistically in five-second clips, Colour Themes let brand admins lock in how colors combine across a Brand Kit while teams still switch between presets, a Presenter Notes Generator writes speaker notes with time estimates from a prompt, and Custom Short Links (canva.link/my-brand-deck) shipped for sharing campaigns without a third-party shortener.
The same release added native integrations with PayPal, TikTok Ads Manager, HubSpot, and Meta Catalog Data, removing an entire export-and-reupload step from a marketing workflow.
ElevenLabs: Voice that finally sounds like it’s reacting
Eleven v3 became generally available in 2026, supporting 70+ languages, multi-speaker dialogue, and audio tags that control emotion directly in the script, things like [whispers] or [sighs]. ElevenLabs’ own framing is an unusually specific admission from a vendor: “the consistent limitation wasn’t sound quality, it was expressiveness.”
The tradeoff: v3 needs more prompt engineering and has higher latency, so ElevenLabs still recommends v2.5 Turbo for live conversational use. On July 9, 2026, it deprecated legacy models, worth knowing if anything you built is pinned to an older model ID. Pricing held steady, with Starter still at $6 a month.
Synthesia: A steady drip of small upgrades beats one big launch
Synthesia’s changelog shows a different pattern than most vendors here: no big-bang release, just a near-weekly cadence through mid-2026. Motion Graphics Translation landed July 24, letting branded animated infographics translate automatically inside the standard workflow. Dubbing 2.0 shipped as what Synthesia calls “a new model, a new bar for quality.”
A Gemini Omni Flash integration went into public preview June 30 for multimodal video generation, and B-roll AI presets followed July 30. For a marketing team on a deadline, a dozen small releases can matter more than one flagship announcement.
3 AI Tools for Business That Got Sharper in 2026
Compliance, integration depth, and cost efficiency drove business-tool improvements more than raw model quality this year.
HubSpot Breeze: AI moved into the places reps actually work
HubSpot’s July 2026 update ships AI where the friction actually is, not as a separate destination. Breeze now generates a summary and next-action suggestion inside contact search results, and reps can invoke it from the CRM email composer without leaving the record. Assignment emails carry action buttons that open Breeze pre-loaded with context, and on mobile it builds meeting briefings from CRM and deals data automatically.
A Microsoft 365 Copilot integration exposes live deal, ticket, and contact data inside Copilot without indexing it into Microsoft’s systems, a real distinction for teams with data-residency requirements. The Prospecting Agent, previously limited to select accounts, is now available to all paid customers.
Zapier Agents: Automation that can browse, not just trigger
Zapier’s agents, still in beta, act autonomously across 7,000+ connected apps and now include web browsing rather than just app-to-app triggers. Dan Martell’s ranking puts Zapier in S tier, citing agencies building real revenue purely on Zapier automation. Worth flagging: Zapier’s own agent documentation is thinner on 2026 launch dates and pricing than other vendors here, so treat the timeline as approximate.
Google Gemini: Three new models, and one notably missing
On July 21, 2026, Google released three new Gemini models rather than the Gemini 3.5 Pro update many expected. Gemini 3.6 Flash is the workhorse model, with coding and multimodal improvements and a 17% reduction in token usage versus its predecessor, lowering inference cost for anyone building on the API. Gemini 3.5 Flash-Lite targets the cheapest end of deployment.
A third model, Gemini 3.5 Flash Cyber, is fine-tuned for cybersecurity vulnerability detection and limited to governments and vetted partners. Bloomberg reportedly linked the missing 3.5 Pro update to internal performance targets not being met: Google shipped around the gap instead of closing it.
5 AI Software Tools That Got Worse in 2026
Every claim in this section is tied to a dated changelog, pricing page, or independent tracker, not a vague impression. Where the picture is genuinely mixed, I’ve said so.
GitHub Copilot: The bill became the feature
On June 1, 2026, Copilot moved from flat pricing to usage-based billing. Copilot Pro now costs $10 a month, including $15 of AI credits, at 1 credit per $0.01. A single agentic session can burn $30 to $40 in credits, three to four times a Pro subscriber’s monthly allotment, which explains how a $29 bill became $750. GitHub’s CPO reasoned that a short chat question and an hours-long session were previously billed identically, which was unsustainable.
The uncomfortable part: the same company spent two years promoting the autonomous workflows now triggering the biggest bills, before metering them.
Otter.ai: Same price, 80% less runway
Otter cut its Pro plan’s transcription allowance from 6,000 minutes to 1,200 minutes a month, keeping the price at $16.99 ($8.33 billed annually). That’s an 80% cut in usable minutes with no price change. A user transcribing 600 minutes a month now pays roughly $0.83 an hour, worse than several usage-based competitors, and a normal meeting schedule now burns through the allowance in under two weeks.
A separate problem surfaced in April 2026: Google Meet began flagging third-party AI notetaker bots as a security risk, affecting Otter in security-conscious workspaces regardless of what Otter itself does.
Adobe Firefly: The free ride has an expiration date
Firefly’s unlimited generation access for select models ends April 22, 2026, after which all plans revert to standard credit consumption that doesn’t roll over. The bigger issue for anyone using Firefly as an access layer to third-party models: FLUX.2 Pro costs roughly $0.10 per image through Firefly versus about $0.03 direct from Black Forest Labs, a 3.3x markup for the convenience of one interface.
ChatGPT: Same subscription, less predictable answers
Caveat up front: DataNorth still ranks ChatGPT first for complex professional work, citing GPT-5.6 and the ChatGPT Work product.
Both things are true at once. Per reporting from April 2026, OpenAI now routes ordinary queries through a tiered inference system, sending simple queries to lighter models and only complex ones to full GPT-5.x, invisibly. GPT-4o, widely considered the best speed-quality balance, was phased out for GPT-5.x variants, and each update reportedly added more restrictive RLHF safety tuning.
Plainly: the lever that makes ChatGPT cheaper to run at the free and Plus tiers is why casual users describe answers as inconsistent, while the paid Work tier routes around that problem.
Salesforce Agentforce: Four pricing models and counting
Salesforce has run through four pricing approaches for Agentforce since fall 2024: flat $2 per conversation, Flex Credits at $0.10 per action, a hybrid Flex Agreement, and per-user licensing now ranging from $125 to roughly $550 monthly by edition. Customers struggled to define a billable “conversation”: short, long, and useless interactions were charged the same. Result: only about 8,000 of 150,000+ customers had adopted Agentforce by May 2025, cost cited as the main blocker, with pricing still being revised into 2026.
AI Tools Comparison Table for 2026
The pattern: AI tools that got better mostly improved efficiency or unlocked a gated feature. And the AI tools that got worse mostly did it through billing, not the product getting dumber.
| Tool | Category | 2026 Direction | Biggest Change | Source |
| Perplexity | Productivity | Better | Comet browser made free | AI Vault |
| Cursor | Productivity | Better | Router cuts model costs by 30 to 50% | DataNorth (cited above) |
| Claude | Productivity | Better | Adoption up to 43% among developers | nxcode.io |
| Notion | Productivity | Mixed | Agent costs down 35 to 50%, AI gated to Business plan | AI Vault |
| Canva | Marketing | Better | Five Magic Studio upgrades in one release | Canva Newsroom |
| ElevenLabs | Marketing | Better | Eleven v3 adds emotional expressiveness | ElevenLabs |
| Synthesia | Marketing | Better | Steady cadence of workflow upgrades | Synthesia |
| HubSpot Breeze | Business | Better | AI embedded directly into rep workflows | Vantage Point |
| Zapier Agents | Business | Better | Agents can now browse the web autonomously | NoCodeFinder |
| Google Gemini | Business | Better | New Flash models cut token cost by 17% | TechCrunch |
| GitHub Copilot | Software | Worse | Usage-based billing spiked some bills 20x+ | Medium |
| Otter.ai | Software | Worse | Pro minutes cut 80%, price unchanged | FelloAI |
| Adobe Firefly | Software | Worse | Unlimited access ends, 3.3x partner-model markup | Future Stack Reviews |
| ChatGPT | Software | Mixed | Tiered routing makes casual use inconsistent | nxcode.io |
| Salesforce Agentforce | Software | Worse | Fourth pricing model since 2024 | Monetizely |
How to Pick AI Tools for Your Stack Without Getting Burned
A few questions worth asking before committing to any tool above:
- Is the price you’re seeing the price you’ll pay?
Usage-based pricing (Copilot, Firefly, Cursor, Salesforce) can swing your bill 10x depending on use, not just tier.
- Does the vendor publish a real changelog?
Canva, ElevenLabs, Synthesia, and HubSpot all have dated, specific release logs. AI tools without one are harder to hold accountable.
- What happens if you leave in 30 days?
Export your data, cancel, and see what breaks. That’s more honest than a features page.
- Is the AI feature gated behind an upgrade you didn’t ask for?
Notion’s shift from a standalone add-on to Business-only access is a pattern worth checking for elsewhere.
Final Thoughts
The clearest pattern across both lists isn’t which company shipped the flashiest feature. It’s that the pricing structure moved faster than product quality did. Four AI tools on the worst list (Copilot, Otter, Firefly, Salesforce) got worse through billing changes, not through the underlying model getting less capable. If 2025 was about which AI tool was smartest, 2026 turned out to be about which one you could actually predict your bill from. Check your own renewal dates before assuming the tool you picked last year still costs what you think it does.
For more info on AI and tech, visit Yaabot.
FAQs
Canva’s May 2026 Magic Studio release is the broadest, with five upgrades and four platform integrations all fitting into one update.
Some AI tools are. In March 2026, Perplexity’s Comet browser became free, while ElevenLabs kept their $6 Starter tier the same. Others, such as Adobe Firefly’s unlimited window, have limited but definite expiration dates, so make sure the “free” is permanent or promotional.
It’s primarily about prices, rather than quality of products. Each of Copilot, Otter.ai, Firefly, and Salesforce has undergone feature reductions that were not dramatic enough to justify their price increases, and each implemented a shift in the billing model or usage limits, meaning they won’t be as free as they once were.
Both are documented. OpenAI’s tiered inference routing system is changing, causing the model used to answer a question to vary depending on the free or Plus tier, and thus resulting in inconsistent outputs. Even though not all rankings are independent, DataNorth still ranks GPT-5.6 and the Work product first for complex tasks involving professional work.
With so many vendors adjusting prices mid-year (Salesforce was the one that made four changes since 2024 and Copilot mid-2026, etc.), monitoring your bill every quarter is better than annually.
