Upon the launch of its latest iteration, the 3.5 Flash model at I/O 2026, Google underscored its enhanced coding capabilities—emphasising the swift deployment of AI agents endowed with intelligence comparable to OpenAI’s renowned GPT-5.5 model.
To evaluate these assertions rigorously, I utilised 3.5 Flash within Google’s revamped Antigravity coding application, continuing the refinement of my Warframe build calculator.
The speed is truly exhilarating; the ability to partition tasks across multiple agents is impressive. Yet, persistent inaccuracies and an inability to adhere to directives mar the experience. Here’s why I prefer ChatGPT for my coding endeavours.
3.5 Flash Trades Accuracy for Raw Speed
The name 3.5 Flash is aptly chosen. Its velocity is startling; I often found it hard to believe tasks were completed within the promised timeframe.
My objective was to enhance my Warframe build calculator, enabling it to accommodate weapon builds along with character profiles.
This necessitated the compilation of an extensive database encompassing every weapon available in the game, numbering in the hundreds.
Each weapon features a plethora of attributes—including critical chance, damage, damage types, fire rate, among others—that need incorporation.
Everything announced at Google I/O 2026 in 13 minutes
When I called upon 3.5 Flash to compile this database, it swiftly generated a script to glean the necessary information online, completing the task in a mere three minutes.
Comparatively, employing ChatGPT and Claude for similar databases had consumed significantly more time.
The usage metrics indicated no remarkable depletion in my allowances, whereas previous attempts with ChatGPT and Claude exerted a far greater toll.
However, did 3.5 Flash excel in crafting my database? Unfortunately, the answer is no. Flash’s propensity to disregard instructions became immediately evident.
I had clearly stated the requirement for it to verify all data against two distinct sources, providing a hierarchy to delineate between credible and less reliable references.
Nevertheless, it generated two URLs for each entry without actually consulting both sources, contravening my explicit directives.
I encountered similar sourcing challenges with ChatGPT and Claude previously. In a bid to rectify this, I instructed 3.5 Flash to access each weapon’s relevant page on the official Warframe wiki, aiming to cross-reference the database contents, not only for inaccuracies but also for any overlooked critical information.
After a minute, Flash declared its task complete. However, the idea of accessing hundreds of web pages in mere seconds felt implausible.
Upon reviewing the resultant markdown file, it became clear that Flash had consulted only a handful of pages and primarily relied on the same script designed for initial extraction.
This approach does not allow for the detection of potentially valuable data points beyond its programmed limits.
Recurring issues—errors, omissions, and stubborn defiance of simple instructions—characterise my encounters with 3.5 Flash.
During my audit of the database, I frequently had to reiterate prompts, as Flash could identify only a smattering of minor issues in one go.
When I attempted to integrate the database into my application to implement the weapon-building functionality, Flash worked for a minute or two, inadvertently damaged my app, and proclaimed the task complete. In summary, 3.5 Flash’s core intelligence falls short of the sophistication exhibited by GPT-5.5 or Opus 4.7.
Agentic Workflows Are Impressive—Until They Aren’t
I appreciate 3.5 Flash’s conceptualisation of AI agents. Visualise Flash as a manager coordinating employees. It deftly deploys agents across various components of a prompt, managing their tasks efficiently.
This parallel processing paradigm is advantageous; while one can convert other large language models into discrete agents via third-party software, such capabilities are not always as readily accessible as they are with Flash in Google’s Antigravity platform.
Regrettably, the agents in Flash do not significantly ameliorate the broader issues I’ve highlighted. During attempts to allocate agents for database creation or to devise an implementation strategy for the integration, errors persisted.
Even while using agents, the errors compounded, with speed diminishing the quality of outputs. The efficiency of condensing a multitude of prompts into a singular interaction is notable, yet it does not counterbalance the frequency of mistakes.
Ultimately, my anticipation for Google’s forthcoming 3.5 Pro model remains high. Theoretically, with a more advanced intelligence guiding agents designed for complex tasks, this functionality could truly excel.
The viability of this vision remains uncertain, but one need not wait extensively, as 3.5 Pro is slated for release in June.
Antigravity Lags Behind Competing Coding Environments
My exploration of 3.5 Flash occurred within Google’s updated Antigravity app. Antigravity positions itself as Google’s answer to Claude Code or OpenAI’s Codex.
It facilitates the use of natural language prompts to accomplish coding tasks, yet it falls short in providing numerous quality-of-life enhancements present in Claude Code and Codex.
For instance, Antigravity fails to display the fullness of your context window within a specific dialogue.
Both Claude Code and Codex offer this feature, allowing users to ascertain when to initiate a new session.
Large context windows often lead to more errors and complications as users approach their limits, necessitating vigilance in monitoring these thresholds.
Antigravity restricts access to usage displays to its settings page, whereas information in Claude Code and Codex is readily available during conversations.
Moreover, it convolutes the data into an awkward array of five distinct bars instead of a straightforward percentage (although the percentage is visible when hovering over the bars).
The bars themselves can be inconsistent; I once reached my usage limit only for the display to inaccurately suggest I had remaining capacity.
I also encountered difficulties when attempting to open my app in Antigravity’s sidebar. While it performed correctly in Chrome and switching tabs was unproblematic, Claude Code and Codex not only facilitate sidebar access but also allow seamless viewing of apps in both desktop and mobile formats.
These shortcomings do not reflect the raw output capabilities of Antigravity or 3.5 Flash, but they do evoke disappointment at Google’s failure to elevate Antigravity closer to its competitive coding counterparts.
Don’t Rush to Switch to 3.5 Flash—Yet
It is imperative to acknowledge that I am a hobbyist coder rather than a professional programmer, so your requirements may diverge from mine.
Nonetheless, 3.5 Flash has not provided sufficient motivation for me to prioritise its use over offerings from Anthropic or OpenAI.
Furthermore, if the absolute pinnacle of AI is not a prerequisite for your coding objectives, a more economical model from DeepSeek could serve you better than 3.5 Flash.
Despite my reservations, I remain vigilant regarding Gemini’s evolving coding capabilities, particularly as 3.5 Pro emerges on the horizon.
Source link: Pcmag.com.
