(Ruben Circelli/PCMag; Anthropic/Embark Studios)
No one who knows me would dare accuse me of being patient, especially when it comes to gaming. (See the apps I made for Warframe and Magic: The Gathering Arena.) But hey, I just always want to be as efficient and expedient as possible. Competitive first-person shooters, like The Finals, are traditionally things you really need to spend time with to succeed at, but that doesn’t mean you can’t try to speed up that process. So, I used Anthropic’s latest complex reasoning model, Opus 5, to build an app that helps me do just that through drills, simulations, and more. Here’s how well Opus 5 worked (and didn’t) for me, as a longtime GPT-5.6 fan.
Introducing FISH: The Finals Interactive Study Helper
If you haven’t played The Finals, it’s a much more chaotic and dynamic game than many other first-person arcade shooters, such as Battlefield or Call of Duty. It relies heavily on creative, in-the-moment decision-making. My Finals Interactive Study Helper (FISH) app thus exists to help with strategy: It generates random game states for me to judge the correct play, and it trains me to memorize gameplay elements, such as specialization icons that briefly pop up beside player names when a game starts. To be clear, FISH is not for cheating. It doesn’t interact with The Finals game files or even run at the same time.
FISH can comprehensively simulate games, including gadgets equipped, player states (alive, dead, and time to respawn), scoreboards, respawn coins available, team compositions, team wipe costs, weapons equipped, and more at different phases and rounds, all within the context of a map that plots objectives, players, and statues. Then, FISH provides 'answers' for what the correct play might be depending on the state it generated.
My app provides two types of answers: rulebook and pro strategy. Rulebook answers boil down to basic advice grounded in a comprehensive understanding of The Finals' rules and mechanics. For example, if it’s late in the first round and your team is currently in third place when there’s only one cashout (the primary objective in The Finals) active, the rulebook answer would be to steal that cashout. Pro strategy answers, however, give better-informed, more specific advice, but more on how those work later.
On the memorization front, I built FISH to drill me on the game's gadgets, interactive items, specializations, maps, and map variants. In general, this means presenting me with three random images of various things without corresponding descriptions, names, and stats so I can quiz myself. Then, I can reveal the information to see what I actually know. For maps, FISH helps me to remember their layouts or which modifiers each map can roll, among other things.
Coding With Opus 5: The Great, the Good, and the Limit-Draining
I built FISH with Opus 5 and Claude’s Max 5x plan ($100 per month). After my experience with Fable 5, one of my biggest takeaways using the latest version of Opus is just how efficient it is. I can send far more prompts compared with Fable before I hit my five-hour limit. However, the existence of five-hour limits makes using Claude infinitely more frustrating than ChatGPT, which now opts for weekly limits.
Of course, Opus 5 can still blow through a five-hour limit incredibly quickly when spinning up agents. You can outline a policy for agents in Claude’s memories, but by default, Claude might set up 10 or 20 agents for a task, exhausting your five-hour limit in minutes. This was an issue when I was auditing FISH. By the time I checked back in on my initial task a few minutes later to see how the work was going, I had nearly exhausted my five-hour limit. This is frustrating, especially because agent failures oftentimes necessitate redoing work.
However, sometimes Opus 5 sets up just one or two agents that run for 20 or 30 minutes. Then another one or two agents run for the same amount of time, working through the entirety of a five-hour window without ever exhausting the limit. I noticed this happen a few times when debugging FISH and refining its underlying parsers. In general, I prefer ChatGPT’s default three-agent-max setup, especially considering the occasionally uneven performance of Opus’ agents.
For whatever reason, Opus’ agents struggle to complete certain tasks, such as searching the web. For example, FISH required a ton of information on The Finals’ items and mechanics, which is available on its official website or wiki. However, Opus’ agents routinely failed to extract the information from these easily accessible web pages, which required me to provide links and quotes. During my time with GPT-5.6’s agents, they didn’t have the same issue.
Opus 5 also consistently has a problem following instructions in its memories. For instance, I added a memory with the instruction to be concise, as Claude tends to be verbose and fill files with meta-commentary. Nonetheless, Opus 5 still routinely did so. As per another memory I set up, Opus 5 was to stop working when it filled up 80% of its context window and update its corresponding handoff document for use in a fresh session. However, it kept stopping work randomly, not knowing how full its context window really was.
More complex tasks don't seem to be as problematic. It only took a few prompts to set up a model to simulate games in The Finals, and it was similarly easy to create a pannable, zoomable map to display simulated game states. What was most impressive about Opus 5 was how easily it created a system to give informed advice on what to do in randomly generated game states.
Feeding AI YouTube Guides: Turning Pro Gameplay Into Strategy Drills
As mentioned above, the games FISH simulates come with answers in the form of advice on how to respond to a given situation. Giving advice based on the basic mechanics and rules of the game is easy, but creating a system to dynamically dole out sophisticated advice for randomly generated game states is much more complicated. Still, Opus 5 made building that system surprisingly easy, even if the result is imperfect.
I started by extracting transcripts of hundreds of YouTube videos on playing The Finals from experienced players. Then, I asked Opus 5 to analyze these transcripts to create a strategy bible. Dozens of agents and millions of tokens later, I had my bible, and Opus 5 was able to distill that bible and its corresponding citations into principles it could then tag relevant game states with as its advice. I can’t overstate how simple all of this was, even if it took Opus 5 a while to crunch through all the data.
Let’s use a random game state I generated with FISH as an example. In this scenario, it’s the final round, my team is ahead, and there’s a cashout nearby controlled by the enemy team. The rulebook's advice would be to push the cashout. However, the pro strategy advice goes further. It also advises a fight, but notes that now is the time for my light teammate to aggressively engage the enemy team, since they’re down a member in the simulation. It further advises my light teammate to keep their equipped glitch grenade at the ready to deal with the enemy team’s equipped gateway gadget.
Currently, FISH has around 70 principles that apply to a variety of situations, so every piece of advice isn’t the same, nor is it obvious given the mechanics and rules of the game. In fact, its rulebook and pro strategy advice often differ. Nonetheless, I still find a lot of value in the more nuanced strategy advice.
Does It Work? How FISH Actually Impacts My Gameplay
Building the app helped me better understand the mechanics and rules of The Finals. I can now more readily recall the wide range of gadgets, interactive items, maps, map modifiers, and specializations in the game, and I am quicker to analyze and respond to situations. FISH routinely reminds me that sometimes wiping as a team and respawning saves time over trying to revive my teammates. It helps me remember that placing goo a level below a cashbox can be an effective way to stop an RPG blast from an enemy team dropping it out of reach. It taught me the difference in playing for a steal on a cashout versus trying to wipe the enemy team and then going for a steal.
For as many videos on The Finals as I watch, it’s hard to remember each discrete piece of advice and apply it to relevant situations in the moment. FISH makes all that easier by reinforcing my understanding of the underlying principles and helping me to put them into action later.
With only a few dozen hours in The Finals so far, I'm still just a total novice, which is to be expected. However, in just a week of using FISH, I’ve learned a lot, and I’ve enjoyed getting better. It's still going to take a lot of time and games to see major improvement, but this project still feels worth the effort.
The Bottom Line: Opus 5 Is Closing the Gap With GPT-5.6
Whereas my comparisons between GPT-5.5 and Opus 4.6, or Fable and GPT-5.6, came out clearly in favor of OpenAI’s models, Opus 5 is genuinely worth consideration for its efficiency gains and a competitive level of intelligence with Fable 5. For my coding projects, including FISH, GPT-5.6 still tends to give me the best results, but each project is different and may not be as well-suited to OpenAI’s models. Whatever you want to build, I recommend at least giving Opus 5 a shot.
About Our Expert
I’ve been writing about consumer technology and video games for over a decade at a variety of publications, including Destructoid, GamesRadar+, Lifewire, PCGamesN, Trusted Reviews, and What Hi-Fi?, among many others. At PCMag, I review AI and productivity software—everything from chatbots to to-do list apps. In my free time, I’m likely cooking something, playing a game, or tinkering with my computer.
- I Can't Code, But I've Made Five Game Mods With AI. Here's How
- AI Built a Photoshop Clone in 24 Hours. Is This Piracy, Plagiarism, or Progress?
- AI Has Ushered In the Era of Personalized Software, and Honestly? I’m Excited
- Airtable Review: Powerful, Flexible Work Management Powered by Deep AI Integration
- I Let 3 AI Agents Take Over My Life. Only 1 Didn't Fail
- More from Ruben Circelli
